swarm_sdk 2.0.0 → 2.0.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 1414a5d98971c8dde0d38481f4510a50770821a667bdbb909cc795538c40fc8a
-  data.tar.gz: fae3e5e52be1d0bb2e7879ea53a57e9b65d3de2e2d77cab42f03c5982ea37ff3
+  metadata.gz: f2cedd1197d0fea1e52b40531ecb6d6f030524436c3164c109bce37da1db883b
+  data.tar.gz: ac77a97bd54374a9049de888e0dfc4a1b1c99d5a4fdc2c01fabe20e54a7cd83b
 SHA512:
-  metadata.gz: 3a02be54f9369614496919d1642d032630c766ed22ceaffed2a31e5400122e4716413ffe1a1d3d8b1b1ce96c27227183224250a4e03a20b3c6763b5fc7f64fc1
-  data.tar.gz: beba552988fe6a545d7e8e5b638591a6890c6e07699f3317c819521c5bb68d02d99d5489de2889aeaf92da3d5097d4320223bed2fae3bd11e5d718fc2c06c1f1
+  metadata.gz: 9d1dfdd64d1247e15a3fea9ce56b0d28657e85a5f9d706d1a2ce9c72700226cf443dcdc691596710d2e948816a3fb5d93fe1520f3bc798df78a2b6a91e9a2085
+  data.tar.gz: b54ae2918f3cb04eb2a49a80faecc2aac1a878cd14828cc3287705eae22ba4da09780c8e38a54c5d048cf4eb1fec75e523dc609192758f6ad60294467d68df01

data/lib/swarm_sdk/agent/builder.rb

@@ -163,13 +163,14 @@ module SwarmSDK
         @description = text
       end
 
-      # Add tools
+      # Set or add tools
       #
       # Uses Set internally to automatically deduplicate tool names across multiple calls.
       # This allows calling tools() multiple times without worrying about duplicates.
       #
       # @param tool_names [Array<Symbol>] Tool names to add
       # @param include_default [Boolean] Whether to include default tools (Read, Grep, etc.)
+      # @param replace [Boolean] If true, replaces existing tools instead of merging (default: false)
       #
       # @example Basic usage with defaults
       #   tools :Grep, :Read  # include_default: true is implicit
@@ -181,7 +182,11 @@ module SwarmSDK
       #   tools :Read
       #   tools :Write, :Edit  # @tools now contains Set[:Read, :Write, :Edit]
       #   tools :Read          # Still Set[:Read, :Write, :Edit] - no duplicate
-      def tools(*tool_names, include_default: true)
+      #
+      # @example Replace tools (for markdown overrides)
+      #   tools :Read, :Write, replace: true  # Replaces all existing tools
+      def tools(*tool_names, include_default: true, replace: false)
+        @tools = Set.new if replace
         @tools.merge(tool_names.map(&:to_sym))
         @include_default_tools = include_default
       end

data/lib/swarm_sdk/agent/definition.rb

@@ -73,12 +73,9 @@ module SwarmSDK
         @timeout = config[:timeout] || DEFAULT_TIMEOUT
         @bypass_permissions = config[:bypass_permissions] || false
         @max_concurrent_tools = config[:max_concurrent_tools]
-        # Default to true when base_url is set, false otherwise (unless explicitly specified)
-        @assume_model_exists = if config.key?(:assume_model_exists)
-          config[:assume_model_exists]
-        else
-          (base_url ? true : false)
-        end
+        # Always assume model exists - SwarmSDK validates models separately using models.json
+        # This prevents RubyLLM from trying to validate models in its registry
+        @assume_model_exists = true
 
         # include_default_tools defaults to true if not specified
         @include_default_tools = config.key?(:include_default_tools) ? config[:include_default_tools] : true
@@ -168,45 +165,53 @@ module SwarmSDK
 
       private
 
-      # Validate that model exists in RubyLLM registry
+      # Validate that model exists in SwarmSDK's model registry
+      #
+      # Uses SwarmSDK's static models.json instead of RubyLLM's dynamic registry.
+      # This provides stable, offline model validation without network calls.
+      #
+      # Process:
+      # 1. Try to find model directly in models.json
+      # 2. If not found, try to resolve as alias and find again
+      # 3. If still not found, return warning with suggestions
       #
       # @return [Hash, nil] Warning hash if model not found, nil otherwise
       def validate_model
-        # Try to find model in registry (searches all providers)
-        RubyLLM.models.find(@model)
-        nil # Model exists
+        # Try direct lookup first
+        model_data = SwarmSDK::Models.all.find { |m| (m["id"] || m[:id]) == @model }
+
+        # If not found, try alias resolution
+        unless model_data
+          resolved_id = SwarmSDK::Models.resolve_alias(@model)
+          # Only search again if alias was different
+          if resolved_id != @model
+            model_data = SwarmSDK::Models.all.find { |m| (m["id"] || m[:id]) == resolved_id }
+          end
+        end
+
+        if model_data
+          nil # Model exists (either directly or via alias)
+        else
+          # Model not found - return warning with suggestions
+          {
+            type: :model_not_found,
+            agent: @name,
+            model: @model,
+            error_message: "Unknown model: #{@model}",
+            suggestions: SwarmSDK::Models.suggest_similar(@model),
+          }
+        end
       rescue StandardError => e
-        # Model not found - return warning with suggestions
+        # Return warning on error
         {
           type: :model_not_found,
           agent: @name,
           model: @model,
           error_message: e.message,
-          suggestions: suggest_similar_models,
+          suggestions: [],
         }
       end
 
-      # Suggest similar models when a model is not found
-      #
-      # @return [Array<Hash>] Up to 3 similar models with their info
-      def suggest_similar_models
-        normalized_query = @model.to_s.downcase.gsub(/[.\-_]/, "")
-
-        RubyLLM.models.all.select do |model_info|
-          normalized_id = model_info.id.downcase.gsub(/[.\-_]/, "")
-          normalized_id.include?(normalized_query) ||
-            model_info.name&.downcase&.gsub(/[.\-_]/, "")&.include?(normalized_query)
-        end.first(3).map do |model_info|
-          {
-            id: model_info.id,
-            name: model_info.name,
-            context_window: model_info.context_window,
-          }
-        end
-      rescue StandardError
-        []
-      end
-
       def build_full_system_prompt(custom_prompt)
         # If coding_agent is false (default), return custom prompt with optional TODO/Scratchpad info
         # If coding_agent is true, include full base prompt for coding tasks
@@ -253,9 +258,27 @@ module SwarmSDK
 
       def render_non_coding_base_prompt
         # Simplified base prompt for non-coding agents
-        # Only includes TODO and Scratchpad tool information
+        # Includes environment info, TODO, and Scratchpad tool information
         # Does not steer towards coding tasks
+        cwd = @directory || Dir.pwd
+        platform = RUBY_PLATFORM
+        os_version = begin
+          %x(uname -sr 2>/dev/null).strip
+        rescue
+          RUBY_PLATFORM
+        end
+        date = Time.now.strftime("%Y-%m-%d")
+
         <<~PROMPT.strip
+          # Environment
+
+          <env>
+          Working directory: #{cwd}
+          Platform: #{platform}
+          OS Version: #{os_version}
+          Today's date: #{date}
+          </env>
+
           # Task Management
 
           You have access to the TodoWrite tool to help you manage and plan tasks. Use this tool to track your progress and give visibility into your work.

data/lib/swarm_sdk/claude_code_agent_adapter.rb

@@ -0,0 +1,205 @@
+# frozen_string_literal: true
+
+module SwarmSDK
+  # Adapter for converting Claude Code agent markdown files to SwarmSDK format
+  #
+  # Claude Code agent files use a different syntax and conventions than SwarmSDK:
+  # - Tools are comma-separated strings instead of arrays
+  # - Model shortcuts like 'sonnet', 'opus', 'haiku' instead of full model IDs
+  # - Tool permissions like 'Write(src/**)' instead of SwarmSDK's permission system
+  # - Required 'name' field in frontmatter
+  #
+  # This adapter:
+  # - Detects Claude Code format by checking frontmatter markers
+  # - Converts tools from comma-separated strings to arrays
+  # - Maps model shortcuts to canonical model IDs
+  # - Strips unsupported tool permission syntax with warnings
+  # - Sets coding_agent: true by default
+  # - Warns about unsupported fields
+  #
+  # @example Parse a Claude Code agent file
+  #   content = File.read('.claude/agents/reviewer.md')
+  #   config = ClaudeCodeAgentAdapter.parse(content, :reviewer)
+  #   agent = Agent::Definition.new(:reviewer, config)
+  #
+  class ClaudeCodeAgentAdapter
+    # Fields supported in Claude Code agent frontmatter
+    SUPPORTED_FIELDS = ["name", "description", "tools", "model"].freeze
+
+    # SwarmSDK documentation URL for reference
+    SWARM_SDK_DOCS_URL = "https://github.com/parruda/claude-swarm/blob/main/docs/v2/README.md"
+
+    # Pattern to detect tool permission syntax like Write(src/**)
+    TOOL_PERMISSION_PATTERN = /^([A-Za-z_]+)\([^)]+\)$/
+
+    class << self
+      # Detect if content appears to be in Claude Code agent format
+      #
+      # Detection is based on tools field type:
+      # - Claude Code: tools is a comma-separated string (e.g., "Read, Write, Bash")
+      # - SwarmSDK: tools is an array (e.g., [Read, Write, Bash])
+      #
+      # Note: The 'name' field alone is not sufficient since SwarmSDK also supports it
+      #
+      # @param content [String] Markdown content with YAML frontmatter
+      # @return [Boolean] true if content appears to be Claude Code format
+      def claude_code_format?(content)
+        return false unless content =~ /\A---\s*\n(.*?)\n---\s*\n/m
+
+        frontmatter_yaml = Regexp.last_match(1)
+        frontmatter = YAML.safe_load(frontmatter_yaml, permitted_classes: [Symbol], aliases: true)
+
+        return false unless frontmatter.is_a?(Hash)
+
+        # Only detect as Claude Code if tools field is a comma-separated string
+        # This is the most reliable indicator since SwarmSDK always uses arrays
+        frontmatter.key?("tools") && frontmatter["tools"].is_a?(String)
+      rescue Psych::SyntaxError
+        false
+      end
+
+      # Parse Claude Code agent markdown and convert to SwarmSDK format
+      #
+      # @param content [String] Markdown content with YAML frontmatter
+      # @param agent_name [Symbol, String] Name of the agent
+      # @param inherit_model [String, nil] Model to use when frontmatter has 'inherit'
+      # @return [Hash] Configuration hash suitable for Agent::Definition.new
+      # @raise [ConfigurationError] if content format is invalid
+      def parse(content, agent_name, inherit_model: nil)
+        new(inherit_model: inherit_model).parse(content, agent_name)
+      end
+    end
+
+    # Initialize adapter with optional context
+    #
+    # @param inherit_model [String, nil] Model to use when frontmatter has 'inherit'
+    def initialize(inherit_model: nil)
+      @inherit_model = inherit_model
+      @warnings = []
+    end
+
+    # Parse Claude Code agent content
+    #
+    # @param content [String] Markdown content with YAML frontmatter
+    # @param agent_name [Symbol, String] Name of the agent
+    # @return [Hash] Configuration hash for Agent::Definition
+    # @raise [ConfigurationError] if format is invalid
+    def parse(content, agent_name)
+      unless content =~ /\A---\s*\n(.*?)\n---\s*\n(.*)\z/m
+        raise ConfigurationError, "Invalid Claude Code agent format. Expected YAML frontmatter followed by prompt content."
+      end
+
+      frontmatter_yaml = Regexp.last_match(1)
+      prompt_content = Regexp.last_match(2).strip
+
+      frontmatter = YAML.safe_load(frontmatter_yaml, permitted_classes: [Symbol], aliases: true)
+
+      unless frontmatter.is_a?(Hash)
+        raise ConfigurationError, "Invalid frontmatter format in Claude Code agent file"
+      end
+
+      config = build_config(frontmatter, prompt_content, agent_name)
+      emit_warnings(agent_name)
+      config
+    end
+
+    private
+
+    # Build SwarmSDK configuration from Claude Code frontmatter
+    def build_config(frontmatter, prompt_content, agent_name)
+      warn_unknown_fields(frontmatter)
+
+      config = {
+        description: frontmatter["description"],
+        system_prompt: prompt_content,
+        coding_agent: true, # Default for Claude Code agents
+      }
+
+      # Parse tools if present
+      if frontmatter["tools"]
+        config[:tools] = parse_tools(frontmatter["tools"])
+      end
+
+      # Parse model if present
+      if frontmatter["model"]
+        config[:model] = resolve_model(frontmatter["model"])
+      end
+
+      config
+    end
+
+    # Parse tools field - handles both comma-separated string and array
+    #
+    # @param tools_field [String, Array] Tools from frontmatter
+    # @return [Array<String>] Array of tool names
+    def parse_tools(tools_field)
+      tools_array = if tools_field.is_a?(String)
+        tools_field.split(",").map(&:strip)
+      else
+        Array(tools_field).map(&:to_s)
+      end
+
+      # Clean tool permissions and collect warnings
+      tools_array.map { |tool| clean_tool_permissions(tool) }.compact
+    end
+
+    # Strip tool permission syntax and warn if detected
+    #
+    # @param tool_string [String] Tool name, possibly with permissions like 'Write(src/**)'
+    # @return [String, nil] Clean tool name, or nil if invalid
+    def clean_tool_permissions(tool_string)
+      if tool_string =~ TOOL_PERMISSION_PATTERN
+        tool_name = Regexp.last_match(1)
+        @warnings << "Tool permission syntax '#{tool_string}' detected in agent file. SwarmSDK supports permissions but uses different syntax. Using '#{tool_name}' without restrictions for now. See SwarmSDK documentation for permission configuration: #{SWARM_SDK_DOCS_URL}"
+        tool_name
+      else
+        tool_string
+      end
+    end
+
+    # Resolve model shortcuts to canonical model IDs
+    #
+    # Uses SwarmSDK::Models.resolve_alias to map shortcuts like 'sonnet'
+    # to the latest model IDs from model_aliases.json.
+    #
+    # @param model_field [String] Model from frontmatter
+    # @return [String, Symbol] Canonical model ID or :inherit symbol
+    def resolve_model(model_field)
+      model_str = model_field.to_s.strip
+
+      # Handle 'inherit' keyword
+      return :inherit if model_str == "inherit"
+
+      # Resolve using SwarmSDK model aliases
+      # This maps 'sonnet' → 'claude-sonnet-4-5-20250929', etc.
+      SwarmSDK::Models.resolve_alias(model_str)
+    end
+
+    # Warn about unknown frontmatter fields
+    def warn_unknown_fields(frontmatter)
+      unknown_fields = frontmatter.keys - SUPPORTED_FIELDS
+
+      unknown_fields.each do |field|
+        @warnings << case field
+        when "hooks"
+          "Hooks configuration detected in agent frontmatter. SwarmSDK handles hooks at the swarm level. See: #{SWARM_SDK_DOCS_URL}"
+        else
+          "Unknown field '#{field}' in Claude Code agent file. Ignoring. Supported fields: #{SUPPORTED_FIELDS.join(", ")}"
+        end
+      end
+    end
+
+    # Emit all collected warnings via LogCollector
+    def emit_warnings(agent_name)
+      return if @warnings.empty?
+
+      @warnings.each do |warning|
+        LogCollector.emit(
+          type: "claude_code_conversion_warning",
+          agent: agent_name,
+          message: warning,
+        )
+      end
+    end
+  end
+end

data/lib/swarm_sdk/configuration.rb

@@ -238,8 +238,10 @@ module SwarmSDK
       # Parse markdown and merge with YAML config
       agent_def_from_file = MarkdownParser.parse(content, name)
 
-      # Merge: markdown file overrides merged_config for fields it defines
-      final_config = merged_config.merge(agent_def_from_file.to_h.compact)
+      # Merge: YAML config overrides markdown file (YAML takes precedence)
+      # This allows YAML to override any settings from the markdown file
+      final_config = agent_def_from_file.to_h.compact.merge(merged_config.compact)
+
       Agent::Definition.new(name, final_config)
     rescue StandardError => e
       raise ConfigurationError, "Error loading agent '#{name}' from file '#{file_path}': #{e.message}"

data/lib/swarm_sdk/markdown_parser.rb

@@ -1,12 +1,41 @@
 # frozen_string_literal: true
 
 module SwarmSDK
+  # Parser for agent markdown files with YAML frontmatter
+  #
+  # Supports two formats:
+  # 1. SwarmSDK format - YAML frontmatter with array-based tools
+  # 2. Claude Code format - Detected and converted via ClaudeCodeAgentAdapter
+  #
+  # Format detection is automatic based on frontmatter structure.
   class MarkdownParser
     FRONTMATTER_PATTERN = /\A---\s*\n(.*?)\n---\s*\n(.*)\z/m
 
     class << self
+      # Parse markdown content into an Agent::Definition
+      #
+      # Automatically detects format (SwarmSDK or Claude Code) and routes
+      # to appropriate parser.
+      #
+      # @param content [String] Markdown content with YAML frontmatter
+      # @param agent_name [Symbol, String, nil] Name of the agent
+      # @return [Agent::Definition] Parsed agent definition
+      # @raise [ConfigurationError] if format is invalid
       def parse(content, agent_name = nil)
-        new(content, agent_name).parse
+        # Detect Claude Code format and route to adapter
+        if ClaudeCodeAgentAdapter.claude_code_format?(content)
+          config = ClaudeCodeAgentAdapter.parse(content, agent_name)
+          # For Claude Code format, agent_name parameter is required since
+          # the 'name' field in frontmatter is Claude Code specific and not used
+          unless agent_name
+            raise ConfigurationError, "Agent name must be provided when parsing Claude Code format"
+          end
+
+          Agent::Definition.new(agent_name.to_sym, config)
+        else
+          # Use standard SwarmSDK format parsing
+          new(content, agent_name).parse
+        end
       end
     end
 

data/lib/swarm_sdk/model_aliases.json

@@ -0,0 +1,5 @@
+{
+  "sonnet": "claude-sonnet-4-5-20250929",
+  "opus": "claude-opus-4-1-20250805",
+  "haiku": "claude-haiku-4-5-20251001"
+}

data/lib/swarm_sdk/models.json

@@ -0,0 +1 @@
+[{"name":"Claude Sonnet 4.5","id":"claude-sonnet-4-5-20250929","provider":"anthropic","family":"claude-sonnet-4-5","context_window":200000,"max_output_tokens":64000,"modalities":{"input":["image","text"],"output":["text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":3.0,"cached_input_per_million":3.75,"output_per_million":15.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Claude Haiku 4.5","id":"claude-haiku-4-5-20251001","provider":"anthropic","family":"claude-haiku-4-5","context_window":200000,"max_output_tokens":64000,"modalities":{"input":["image","text"],"output":["text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.0,"cached_input_per_million":1.25,"output_per_million":5.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Claude Opus 4.1","id":"claude-opus-4-1-20250805","provider":"anthropic","family":"claude-opus-4-1","context_window":200000,"max_output_tokens":32000,"modalities":{"input":["image","text"],"output":["text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":18.75,"output_per_million":75.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.5 Pro","id":"gemini-2.5-pro","provider":"gemini","family":"gemini-2.5-pro","context_window":1048576,"max_output_tokens":65536,"modalities":{"input":["audio","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":0.125,"output_per_million":10.0},"batch":{"input_per_million":0.625,"output_per_million":5.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.5 Flash","id":"gemini-2.5-flash","provider":"gemini","family":"gemini-2.5-flash","context_window":1048576,"max_output_tokens":65536,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.3,"cached_input_per_million":0.03,"output_per_million":2.5},"batch":{"input_per_million":0.15,"output_per_million":1.25}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.5 Flash-Lite","id":"gemini-2.5-flash-lite","provider":"gemini","family":"gemini-2.5-flash-lite","context_window":1048576,"max_output_tokens":65536,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.3,"cached_input_per_million":0.03,"output_per_million":2.5},"batch":{"input_per_million":0.15,"output_per_million":1.25}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.0 Flash","id":"gemini-2.0-flash","provider":"gemini","family":"gemini-2.0-flash","context_window":1048576,"max_output_tokens":8192,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.0 Flash","id":"gemini-2.0-flash-001","provider":"gemini","family":"gemini-2.0-flash","context_window":1048576,"max_output_tokens":8192,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.0 Flash","id":"gemini-2.0-flash-exp","provider":"gemini","family":"gemini-2.0-flash","context_window":1048576,"max_output_tokens":8192,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.0 Flash-Lite","id":"gemini-2.0-flash-lite","provider":"gemini","family":"gemini-2.0-flash-lite","context_window":1048576,"max_output_tokens":8192,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Gemini 2.0 Flash-Lite","id":"gemini-2.0-flash-lite-001","provider":"gemini","family":"gemini-2.0-flash-lite","context_window":1048576,"max_output_tokens":8192,"modalities":{"input":["audio","image","text"],"output":["text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"deepseek-chat","id":"deepseek-chat","provider":"deepseek","family":"deepseek-chat","context_window":128000,"max_output_tokens":8000,"modalities":{"input":["text"],"output":["text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.28,"cached_input_per_million":0.028,"output_per_million":0.42},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"deepseek-reasoner","id":"deepseek-reasoner","provider":"deepseek","family":"deepseek-reasoner","context_window":null,"max_output_tokens":64000,"modalities":{"input":["text"],"output":["text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5","family":"gpt-5","provider":"openai","id":"gpt-5-2025-08-07","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":0.125,"output_per_million":10.0},"batch":{"input_per_million":0.625,"output_per_million":5.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5","family":"gpt-5","provider":"openai","id":"gpt-5","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":0.125,"output_per_million":10.0},"batch":{"input_per_million":0.625,"output_per_million":5.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 mini","family":"gpt-5-mini","provider":"openai","id":"gpt-5-mini-2025-08-07","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.25,"cached_input_per_million":0.025,"output_per_million":2.0},"batch":{"input_per_million":0.125,"output_per_million":1.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 mini","family":"gpt-5-mini","provider":"openai","id":"gpt-5-mini","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.25,"cached_input_per_million":0.025,"output_per_million":2.0},"batch":{"input_per_million":0.125,"output_per_million":1.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 nano","family":"gpt-5-nano","provider":"openai","id":"gpt-5-nano-2025-08-07","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.05,"cached_input_per_million":0.005,"output_per_million":0.4},"batch":{"input_per_million":0.025,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 nano","family":"gpt-5-nano","provider":"openai","id":"gpt-5-nano","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.05,"cached_input_per_million":0.005,"output_per_million":0.4},"batch":{"input_per_million":0.025,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 pro","family":"gpt-5-pro","provider":"openai","id":"gpt-5-pro-2025-10-06","context_window":400000,"max_output_tokens":272000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":null,"output_per_million":120.0},"batch":{"input_per_million":7.5,"output_per_million":60.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 pro","family":"gpt-5-pro","provider":"openai","id":"gpt-5-pro","context_window":400000,"max_output_tokens":272000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":null,"output_per_million":120.0},"batch":{"input_per_million":7.5,"output_per_million":60.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1","family":"gpt-4.1","provider":"openai","id":"gpt-4.1-2025-04-14","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1","family":"gpt-4.1","provider":"openai","id":"gpt-4.1","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-oss-120b","family":"gpt-oss-120b","provider":"openai","id":"gpt-oss-120b","context_window":131072,"max_output_tokens":131072,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-oss-20b","family":"gpt-oss-20b","provider":"openai","id":"gpt-oss-20b","context_window":131072,"max_output_tokens":131072,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Sora 2","family":"sora-2","provider":"openai","id":"sora-2","context_window":null,"max_output_tokens":null,"modalities":{"input":["image","text"],"output":["audio","embeddings"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Sora 2 Pro","family":"sora-2-pro","provider":"openai","id":"sora-2-pro","context_window":null,"max_output_tokens":null,"modalities":{"input":["image","text"],"output":["audio","embeddings"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-deep-research","family":"o3-deep-research","provider":"openai","id":"o3-deep-research-2025-06-26","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":10.0,"cached_input_per_million":2.5,"output_per_million":40.0},"batch":{"input_per_million":5.0,"output_per_million":20.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-deep-research","family":"o3-deep-research","provider":"openai","id":"o3-deep-research","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":10.0,"cached_input_per_million":2.5,"output_per_million":40.0},"batch":{"input_per_million":5.0,"output_per_million":20.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o4-mini-deep-research","family":"o4-mini-deep-research","provider":"openai","id":"o4-mini-deep-research-2025-06-26","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o4-mini-deep-research","family":"o4-mini-deep-research","provider":"openai","id":"o4-mini-deep-research","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT Image 1","family":"gpt-image-1","provider":"openai","id":"gpt-image-1","context_window":null,"max_output_tokens":null,"modalities":{"input":["image","text"],"output":["embeddings","image"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":5.0,"cached_input_per_million":1.25,"output_per_million":40.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-image-1-mini","family":"gpt-image-1-mini","provider":"openai","id":"gpt-image-1-mini","context_window":null,"max_output_tokens":null,"modalities":{"input":["image","text"],"output":["embeddings","image"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.2,"output_per_million":8.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"DALL·E 3","family":"dall-e-3","provider":"openai","id":"dall-e-3","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["embeddings","image"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini TTS","family":"gpt-4o-mini-tts","provider":"openai","id":"gpt-4o-mini-tts","context_window":2000,"max_output_tokens":null,"modalities":{"input":["text"],"output":["audio","embeddings"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":null,"output_per_million":12.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o Transcribe","family":"gpt-4o-transcribe","provider":"openai","id":"gpt-4o-transcribe","context_window":16000,"max_output_tokens":2000,"modalities":{"input":["audio","text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Transcribe","family":"gpt-4o-mini-transcribe","provider":"openai","id":"gpt-4o-mini-transcribe","context_window":16000,"max_output_tokens":2000,"modalities":{"input":["audio","text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":null,"output_per_million":5.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-realtime","family":"gpt-realtime","provider":"openai","id":"gpt-realtime-2025-08-28","context_window":32000,"max_output_tokens":4096,"modalities":{"input":["audio","image","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":4.0,"cached_input_per_million":0.5,"output_per_million":16.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-realtime","family":"gpt-realtime","provider":"openai","id":"gpt-realtime","context_window":32000,"max_output_tokens":4096,"modalities":{"input":["audio","image","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":4.0,"cached_input_per_million":0.5,"output_per_million":16.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-audio","family":"gpt-audio","provider":"openai","id":"gpt-audio-2025-08-28","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-audio","family":"gpt-audio","provider":"openai","id":"gpt-audio","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-realtime-mini","family":"gpt-realtime-mini","provider":"openai","id":"gpt-realtime-mini-2025-10-06","context_window":32000,"max_output_tokens":4096,"modalities":{"input":["audio","image","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":0.06,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-realtime-mini","family":"gpt-realtime-mini","provider":"openai","id":"gpt-realtime-mini","context_window":32000,"max_output_tokens":4096,"modalities":{"input":["audio","image","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":0.06,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-audio-mini","family":"gpt-audio-mini","provider":"openai","id":"gpt-audio-mini-2025-10-06","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":null,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-audio-mini","family":"gpt-audio-mini","provider":"openai","id":"gpt-audio-mini","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":null,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5 Chat","family":"gpt-5-chat-latest","provider":"openai","id":"gpt-5-chat-latest","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":0.125,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"ChatGPT-4o","family":"chatgpt-4o-latest","provider":"openai","id":"chatgpt-4o-latest","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":5.0,"cached_input_per_million":null,"output_per_million":15.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-5-Codex","family":"gpt-5-codex","provider":"openai","id":"gpt-5-codex","context_window":400000,"max_output_tokens":128000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.25,"cached_input_per_million":0.125,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-pro","family":"o3-pro","provider":"openai","id":"o3-pro-2025-06-10","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":20.0,"cached_input_per_million":null,"output_per_million":80.0},"batch":{"input_per_million":10.0,"output_per_million":40.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-pro","family":"o3-pro","provider":"openai","id":"o3-pro","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":20.0,"cached_input_per_million":null,"output_per_million":80.0},"batch":{"input_per_million":10.0,"output_per_million":40.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3","family":"o3","provider":"openai","id":"o3-2025-04-16","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3","family":"o3","provider":"openai","id":"o3","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":0.5,"output_per_million":8.0},"batch":{"input_per_million":1.0,"output_per_million":4.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o4-mini","family":"o4-mini","provider":"openai","id":"o4-mini-2025-04-16","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.275,"output_per_million":4.4},"batch":{"input_per_million":0.55,"output_per_million":2.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o4-mini","family":"o4-mini","provider":"openai","id":"o4-mini","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.275,"output_per_million":4.4},"batch":{"input_per_million":0.55,"output_per_million":2.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1 mini","family":"gpt-4.1-mini","provider":"openai","id":"gpt-4.1-mini-2025-04-14","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.4,"cached_input_per_million":0.1,"output_per_million":1.6},"batch":{"input_per_million":0.2,"output_per_million":0.8}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1 mini","family":"gpt-4.1-mini","provider":"openai","id":"gpt-4.1-mini","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.4,"cached_input_per_million":0.1,"output_per_million":1.6},"batch":{"input_per_million":0.2,"output_per_million":0.8}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1 nano","family":"gpt-4.1-nano","provider":"openai","id":"gpt-4.1-nano-2025-04-14","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.1 nano","family":"gpt-4.1-nano","provider":"openai","id":"gpt-4.1-nano","context_window":1047576,"max_output_tokens":32768,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":0.025,"output_per_million":0.4},"batch":{"input_per_million":0.05,"output_per_million":0.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1-pro","family":"o1-pro","provider":"openai","id":"o1-pro-2025-03-19","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":150.0,"cached_input_per_million":null,"output_per_million":600.0},"batch":{"input_per_million":75.0,"output_per_million":300.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1-pro","family":"o1-pro","provider":"openai","id":"o1-pro","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":150.0,"cached_input_per_million":null,"output_per_million":600.0},"batch":{"input_per_million":75.0,"output_per_million":300.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"computer-use-preview","family":"computer-use-preview","provider":"openai","id":"computer-use-preview-2025-03-11","context_window":8192,"max_output_tokens":1024,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":3.0,"cached_input_per_million":null,"output_per_million":12.0},"batch":{"input_per_million":1.5,"output_per_million":6.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"computer-use-preview","family":"computer-use-preview","provider":"openai","id":"computer-use-preview","context_window":8192,"max_output_tokens":1024,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":3.0,"cached_input_per_million":null,"output_per_million":12.0},"batch":{"input_per_million":1.5,"output_per_million":6.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Search Preview","family":"gpt-4o-mini-search-preview","provider":"openai","id":"gpt-4o-mini-search-preview-2025-03-11","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":null,"output_per_million":0.6},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Search Preview","family":"gpt-4o-mini-search-preview","provider":"openai","id":"gpt-4o-mini-search-preview","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":null,"output_per_million":0.6},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o Search Preview","family":"gpt-4o-search-preview","provider":"openai","id":"gpt-4o-search-preview-2025-03-11","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o Search Preview","family":"gpt-4o-search-preview","provider":"openai","id":"gpt-4o-search-preview","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.5 Preview (Deprecated)","family":"gpt-4.5-preview","provider":"openai","id":"gpt-4.5-preview-2025-02-27","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":75.0,"cached_input_per_million":37.5,"output_per_million":150.0},"batch":{"input_per_million":37.5,"output_per_million":75.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4.5 Preview (Deprecated)","family":"gpt-4.5-preview","provider":"openai","id":"gpt-4.5-preview","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":75.0,"cached_input_per_million":37.5,"output_per_million":150.0},"batch":{"input_per_million":37.5,"output_per_million":75.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-mini","family":"o3-mini","provider":"openai","id":"o3-mini-2025-01-31","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.55,"output_per_million":4.4},"batch":{"input_per_million":0.55,"output_per_million":2.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o3-mini","family":"o3-mini","provider":"openai","id":"o3-mini","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.55,"output_per_million":4.4},"batch":{"input_per_million":0.55,"output_per_million":2.2}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Audio","family":"gpt-4o-mini-audio-preview","provider":"openai","id":"gpt-4o-mini-audio-preview-2024-12-17","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":null,"output_per_million":0.6},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Audio","family":"gpt-4o-mini-audio-preview","provider":"openai","id":"gpt-4o-mini-audio-preview","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":null,"output_per_million":0.6},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Realtime","family":"gpt-4o-mini-realtime-preview","provider":"openai","id":"gpt-4o-mini-realtime-preview-2024-12-17","context_window":16000,"max_output_tokens":4096,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":0.3,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini Realtime","family":"gpt-4o-mini-realtime-preview","provider":"openai","id":"gpt-4o-mini-realtime-preview","context_window":16000,"max_output_tokens":4096,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.6,"cached_input_per_million":0.3,"output_per_million":2.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1","family":"o1","provider":"openai","id":"o1-2024-12-17","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":7.5,"output_per_million":60.0},"batch":{"input_per_million":7.5,"output_per_million":30.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1","family":"o1","provider":"openai","id":"o1","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":7.5,"output_per_million":60.0},"batch":{"input_per_million":7.5,"output_per_million":30.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"omni-moderation","family":"omni-moderation-latest","provider":"openai","id":"omni-moderation-latest","context_window":null,"max_output_tokens":null,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.0,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1-mini","family":"o1-mini","provider":"openai","id":"o1-mini-2024-09-12","context_window":128000,"max_output_tokens":65536,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.55,"output_per_million":4.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1-mini","family":"o1-mini","provider":"openai","id":"o1-mini","context_window":128000,"max_output_tokens":65536,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":1.1,"cached_input_per_million":0.55,"output_per_million":4.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1 Preview","family":"o1-preview","provider":"openai","id":"o1-preview-2024-09-12","context_window":128000,"max_output_tokens":32768,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":7.5,"output_per_million":60.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"o1 Preview","family":"o1-preview","provider":"openai","id":"o1-preview","context_window":128000,"max_output_tokens":32768,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":7.5,"output_per_million":60.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o","family":"gpt-4o","provider":"openai","id":"gpt-4o-2024-08-06","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":1.25,"output_per_million":10.0},"batch":{"input_per_million":1.25,"output_per_million":5.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o","family":"gpt-4o","provider":"openai","id":"gpt-4o","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":1.25,"output_per_million":10.0},"batch":{"input_per_million":1.25,"output_per_million":5.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o Audio","family":"gpt-4o-audio-preview","provider":"openai","id":"gpt-4o-audio-preview","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini","family":"gpt-4o-mini","provider":"openai","id":"gpt-4o-mini-2024-07-18","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":0.075,"output_per_million":0.6},"batch":{"input_per_million":0.075,"output_per_million":0.3}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o mini","family":"gpt-4o-mini","provider":"openai","id":"gpt-4o-mini","context_window":128000,"max_output_tokens":16384,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.15,"cached_input_per_million":0.075,"output_per_million":0.6},"batch":{"input_per_million":0.075,"output_per_million":0.3}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4o Realtime","family":"gpt-4o-realtime-preview","provider":"openai","id":"gpt-4o-realtime-preview","context_window":32000,"max_output_tokens":4096,"modalities":{"input":["audio","text"],"output":["audio","embeddings","text"]},"capabilities":["function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":5.0,"cached_input_per_million":2.5,"output_per_million":20.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4 Turbo","family":"gpt-4-turbo","provider":"openai","id":"gpt-4-turbo-2024-04-09","context_window":128000,"max_output_tokens":4096,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":10.0,"cached_input_per_million":null,"output_per_million":30.0},"batch":{"input_per_million":5.0,"output_per_million":15.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4 Turbo","family":"gpt-4-turbo","provider":"openai","id":"gpt-4-turbo","context_window":128000,"max_output_tokens":4096,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["batch","function_calling"],"pricing":{"text_tokens":{"standard":{"input_per_million":10.0,"cached_input_per_million":null,"output_per_million":30.0},"batch":{"input_per_million":5.0,"output_per_million":15.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"babbage-002","family":"babbage-002","provider":"openai","id":"babbage-002","context_window":null,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.4,"cached_input_per_million":null,"output_per_million":0.4},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"codex-mini-latest","family":"codex-mini-latest","provider":"openai","id":"codex-mini-latest","context_window":200000,"max_output_tokens":100000,"modalities":{"input":["image","text"],"output":["embeddings","text"]},"capabilities":["function_calling","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":1.5,"cached_input_per_million":0.375,"output_per_million":6.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"DALL·E 2","family":"dall-e-2","provider":"openai","id":"dall-e-2","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["embeddings","image"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"davinci-002","family":"davinci-002","provider":"openai","id":"davinci-002","context_window":null,"max_output_tokens":16384,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":2.0,"cached_input_per_million":null,"output_per_million":2.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-3.5 Turbo","family":"gpt-3.5-turbo","provider":"openai","id":"gpt-3.5-turbo","context_window":16385,"max_output_tokens":4096,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.5,"cached_input_per_million":null,"output_per_million":1.5},"batch":{"input_per_million":0.25,"output_per_million":0.75}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4","family":"gpt-4","provider":"openai","id":"gpt-4-0613","context_window":8192,"max_output_tokens":8192,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":30.0,"cached_input_per_million":null,"output_per_million":60.0},"batch":{"input_per_million":15.0,"output_per_million":30.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4","family":"gpt-4","provider":"openai","id":"gpt-4","context_window":8192,"max_output_tokens":8192,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":30.0,"cached_input_per_million":null,"output_per_million":60.0},"batch":{"input_per_million":15.0,"output_per_million":30.0}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"GPT-4 Turbo Preview","family":"gpt-4-turbo-preview","provider":"openai","id":"gpt-4-turbo-preview","context_window":128000,"max_output_tokens":4096,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":10.0,"cached_input_per_million":null,"output_per_million":30.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"gpt-4o-transcribe-diarize","family":"gpt-4o-transcribe-diarize","provider":"openai","id":"gpt-4o-transcribe-diarize","context_window":16000,"max_output_tokens":2000,"modalities":{"input":["audio","text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":2.5,"cached_input_per_million":null,"output_per_million":10.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"text-embedding-3-large","family":"text-embedding-3-large","provider":"openai","id":"text-embedding-3-large","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.13,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":0.065,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":0.13},"batch":{"input_per_million":0.065}}}},{"name":"text-embedding-3-small","family":"text-embedding-3-small","provider":"openai","id":"text-embedding-3-small","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch","structured_output"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.02,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":0.01,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":0.02},"batch":{"input_per_million":0.01}}}},{"name":"text-embedding-ada-002","family":"text-embedding-ada-002","provider":"openai","id":"text-embedding-ada-002","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":["batch"],"pricing":{"text_tokens":{"standard":{"input_per_million":0.1,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":0.05,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":0.1},"batch":{"input_per_million":0.05}}}},{"name":"text-moderation","family":"text-moderation-latest","provider":"openai","id":"text-moderation-latest","context_window":null,"max_output_tokens":32768,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.0,"cached_input_per_million":null,"output_per_million":0.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"text-moderation-stable","family":"text-moderation-stable","provider":"openai","id":"text-moderation-stable","context_window":null,"max_output_tokens":32768,"modalities":{"input":["text"],"output":["embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.0,"cached_input_per_million":null,"output_per_million":0.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"TTS-1","family":"tts-1","provider":"openai","id":"tts-1","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["audio","embeddings"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":15.0,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"TTS-1 HD","family":"tts-1-hd","provider":"openai","id":"tts-1-hd","context_window":null,"max_output_tokens":null,"modalities":{"input":["text"],"output":["audio","embeddings"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":null,"cached_input_per_million":null,"output_per_million":30.0},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}},{"name":"Whisper","family":"whisper-1","provider":"openai","id":"whisper-1","context_window":null,"max_output_tokens":null,"modalities":{"input":["audio"],"output":["audio","embeddings","text"]},"capabilities":[],"pricing":{"text_tokens":{"standard":{"input_per_million":0.006,"cached_input_per_million":null,"output_per_million":null},"batch":{"input_per_million":null,"output_per_million":null}},"embeddings":{"standard":{"input_per_million":null},"batch":{"input_per_million":null}}}}]

data/lib/swarm_sdk/models.rb

@@ -0,0 +1,120 @@
+# frozen_string_literal: true
+
+module SwarmSDK
+  # Models provides model validation and suggestion functionality
+  #
+  # Uses static JSON files:
+  # - models.json: Curated model list from Parsera
+  # - model_aliases.json: Shortcuts mapping to latest models
+  #
+  # This avoids network calls, API key requirements, and RubyLLM
+  # registry manipulation.
+  #
+  # @example
+  #   model = SwarmSDK::Models.find("claude-sonnet-4-5-20250929")
+  #   model = SwarmSDK::Models.find("sonnet")  # Uses alias
+  #   suggestions = SwarmSDK::Models.suggest_similar("anthropic:claude-sonnet-4-5")
+  class Models
+    MODELS_JSON_PATH = File.expand_path("models.json", __dir__)
+    ALIASES_JSON_PATH = File.expand_path("model_aliases.json", __dir__)
+
+    class << self
+      # Find a model by ID or alias
+      #
+      # @param model_id [String] Model ID or alias to find
+      # @return [Hash, nil] Model data or nil if not found
+      def find(model_id)
+        # Check if it's an alias first
+        resolved_id = resolve_alias(model_id)
+
+        all.find { |m| m["id"] == resolved_id || m[:id] == resolved_id }
+      end
+
+      # Resolve a model alias to full model ID
+      #
+      # @param model_id [String] Model ID or alias
+      # @return [String] Resolved model ID (or original if not an alias)
+      def resolve_alias(model_id)
+        aliases[model_id.to_s] || model_id
+      end
+
+      # Suggest similar models for a given query
+      #
+      # Strips provider prefixes and normalizes for fuzzy matching.
+      #
+      # @param query [String] Model ID to match against
+      # @param limit [Integer] Maximum number of suggestions
+      # @return [Array<Hash>] Up to `limit` similar models
+      def suggest_similar(query, limit: 3)
+        # Strip provider prefix (e.g., "anthropic:claude-sonnet-4-5" → "claude-sonnet-4-5")
+        query_without_prefix = query.to_s.sub(/^[^:]+:/, "")
+        normalized_query = query_without_prefix.downcase.gsub(/[.\-_]/, "")
+
+        matches = all.select do |model|
+          model_id = (model["id"] || model[:id]).to_s
+          model_name = (model["name"] || model[:name]).to_s
+
+          normalized_id = model_id.downcase.gsub(/[.\-_]/, "")
+          normalized_name = model_name.downcase.gsub(/[.\-_]/, "")
+
+          normalized_id.include?(normalized_query) || normalized_name.include?(normalized_query)
+        end.first(limit)
+
+        matches.map do |m|
+          {
+            id: m["id"] || m[:id],
+            name: m["name"] || m[:name],
+            context_window: m["context_window"] || m[:context_window],
+          }
+        end
+      end
+
+      # Get all models
+      #
+      # @return [Array<Hash>] All models from models.json
+      def all
+        @models ||= load_models
+      end
+
+      # Get all aliases
+      #
+      # @return [Hash] Alias mappings
+      def aliases
+        @aliases ||= load_aliases
+      end
+
+      # Reload models and aliases from JSON files
+      #
+      # @return [Array<Hash>] Loaded models
+      def reload!
+        @models = load_models
+        @aliases = load_aliases
+        @models
+      end
+
+      private
+
+      # Load models from JSON file
+      #
+      # @return [Array<Hash>] Models array
+      def load_models
+        JSON.parse(File.read(MODELS_JSON_PATH))
+      rescue StandardError => e
+        # Log error and return empty array
+        RubyLLM.logger.error("Failed to load SwarmSDK models.json: #{e.class} - #{e.message}")
+        []
+      end
+
+      # Load aliases from JSON file
+      #
+      # @return [Hash] Alias mappings
+      def load_aliases
+        JSON.parse(File.read(ALIASES_JSON_PATH))
+      rescue StandardError => e
+        # Log error and return empty hash
+        RubyLLM.logger.debug("Failed to load SwarmSDK model_aliases.json: #{e.class} - #{e.message}")
+        {}
+      end
+    end
+  end
+end

data/lib/swarm_sdk/swarm/builder.rb

@@ -94,16 +94,19 @@ module SwarmSDK
       #     You build APIs.
       #   MD
       def agent(name, content = nil, &block)
-        # Case 1: agent :name, <<~MD (name + markdown content)
-        if content.is_a?(String) && !block_given? && markdown_content?(content)
+        # Case 1: agent :name, <<~MD do ... end (markdown + overrides)
+        if content.is_a?(String) && block_given? && markdown_content?(content)
+          load_agent_from_markdown_with_overrides(content, name, &block)
+        # Case 2: agent :name, <<~MD (markdown only)
+        elsif content.is_a?(String) && !block_given? && markdown_content?(content)
           load_agent_from_markdown(content, name)
-        # Case 2: agent :name do ... end (inline DSL)
+        # Case 3: agent :name do ... end (inline DSL)
         elsif block_given?
           builder = Agent::Builder.new(name)
           builder.instance_eval(&block)
           @agents[name] = builder
         else
-          raise ArgumentError, "Invalid agent definition. Use: agent :name { ... } OR agent :name, <<~MD ... MD"
+          raise ArgumentError, "Invalid agent definition. Use: agent :name { ... } OR agent :name, <<~MD ... MD OR agent :name, <<~MD do ... end"
         end
       end
 
@@ -225,6 +228,76 @@ module SwarmSDK
         @agents[definition.name] = { __file_config__: definition.to_h }
       end
 
+      # Load an agent from markdown content with DSL overrides
+      #
+      # This allows loading from a file and then overriding specific settings:
+      #   agent :reviewer, File.read("reviewer.md") do
+      #     provider :openai
+      #     model "gpt-4o"
+      #   end
+      #
+      # @param content [String] Markdown content with frontmatter
+      # @param name_override [Symbol, nil] Optional name to override frontmatter name
+      # @yield Block with DSL overrides
+      # @return [void]
+      def load_agent_from_markdown_with_overrides(content, name_override = nil, &block)
+        # Parse markdown content first
+        definition = MarkdownParser.parse(content, name_override)
+
+        # Create a builder with the markdown config
+        builder = Agent::Builder.new(definition.name)
+
+        # Apply markdown settings to builder (these become the base)
+        apply_definition_to_builder(builder, definition.to_h)
+
+        # Apply DSL overrides (these override the markdown settings)
+        builder.instance_eval(&block)
+
+        # Store the builder (not file config) so overrides are preserved
+        @agents[definition.name] = builder
+      end
+
+      # Apply agent definition hash to a builder
+      #
+      # @param builder [Agent::Builder] Builder to configure
+      # @param config [Hash] Configuration hash from definition
+      # @return [void]
+      def apply_definition_to_builder(builder, config)
+        builder.description(config[:description]) if config[:description]
+        builder.model(config[:model]) if config[:model]
+        builder.provider(config[:provider]) if config[:provider]
+        builder.base_url(config[:base_url]) if config[:base_url]
+        builder.api_version(config[:api_version]) if config[:api_version]
+        builder.context_window(config[:context_window]) if config[:context_window]
+        builder.system_prompt(config[:system_prompt]) if config[:system_prompt]
+        builder.directory(config[:directory]) if config[:directory]
+        builder.timeout(config[:timeout]) if config[:timeout]
+        builder.parameters(config[:parameters]) if config[:parameters]
+        builder.headers(config[:headers]) if config[:headers]
+        builder.coding_agent(config[:coding_agent]) unless config[:coding_agent].nil?
+        # Don't apply assume_model_exists from markdown - let DSL overrides or auto-enable handle it
+        # builder.assume_model_exists(config[:assume_model_exists]) unless config[:assume_model_exists].nil?
+        builder.bypass_permissions(config[:bypass_permissions]) if config[:bypass_permissions]
+        builder.include_default_tools(config[:include_default_tools]) unless config[:include_default_tools].nil?
+
+        # Add tools from markdown
+        if config[:tools]&.any?
+          # Extract tool names from the tools array (which may be hashes with permissions)
+          tool_names = config[:tools].map do |tool|
+            tool.is_a?(Hash) ? tool[:name] : tool
+          end
+          builder.tools(*tool_names)
+        end
+
+        # Add delegates_to
+        builder.delegates_to(*config[:delegates_to]) if config[:delegates_to]&.any?
+
+        # Add MCP servers
+        config[:mcp_servers]&.each do |server|
+          builder.mcp_server(server[:name], **server.except(:name))
+        end
+      end
+
       # Build a traditional single-swarm execution
       #
       # @return [Swarm] Configured swarm instance

data/lib/swarm_sdk/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module SwarmSDK
-  VERSION = "2.0.0"
+  VERSION = "2.0.2"
 end

data/lib/swarm_sdk.rb

@@ -41,36 +41,6 @@ module SwarmSDK
   class StateError < Error; end
 
   class << self
-    # Refresh RubyLLM model registry silently (without log output)
-    #
-    # By default, RubyLLM.models.refresh! outputs INFO level logs about
-    # fetching models from providers. This method temporarily raises the
-    # log level to suppress those messages, which is useful for CLI tools
-    # that want clean output.
-    #
-    # If model refresh fails (e.g., missing API keys, invalid keys, network
-    # unavailable), the error is silently caught and execution continues
-    # using the bundled models.json. This allows SwarmSDK to work offline
-    # and with dummy keys for local proxies.
-    #
-    # @example
-    #   SwarmSDK.refresh_models_silently
-    #
-    # @return [void]
-    def refresh_models_silently
-      original_level = RubyLLM.logger.level
-      RubyLLM.logger.level = Logger::ERROR
-
-      RubyLLM.models.refresh!
-    rescue StandardError => e
-      # Silently ignore all refresh failures
-      # Models will use bundled models.json instead
-      RubyLLM.logger.debug("Model refresh skipped: #{e.class} - #{e.message}")
-      nil
-    ensure
-      RubyLLM.logger.level = original_level
-    end
-
     # Main entry point for DSL
     def build(&block)
       Swarm::Builder.build(&block)
@@ -124,3 +94,19 @@ RubyLLM.configure do |config|
   config.gpustack_api_base ||= ENV["GPUSTACK_API_BASE"]
   config.gpustack_api_key ||= ENV["GPUSTACK_API_KEY"]
 end
+
+# monkey patch ruby_llm/mcp to add `id` when sending "notifications/initialized" message
+# https://github.com/patvice/ruby_llm-mcp/issues/65
+require "ruby_llm/mcp/notifications/initialize"
+
+module RubyLLM
+  module MCP
+    module Notifications
+      class Initialize
+        def call
+          @coordinator.request(notification_body, add_id: true, wait_for_response: false)
+        end
+      end
+    end
+  end
+end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: swarm_sdk
 version: !ruby/object:Gem::Version
-  version: 2.0.0
+  version: 2.0.2
 platform: ruby
 authors:
 - Paulo Arruda
@@ -86,6 +86,7 @@ files:
 - lib/swarm_sdk/agent/chat/system_reminder_injector.rb
 - lib/swarm_sdk/agent/context.rb
 - lib/swarm_sdk/agent/definition.rb
+- lib/swarm_sdk/claude_code_agent_adapter.rb
 - lib/swarm_sdk/configuration.rb
 - lib/swarm_sdk/context_compactor.rb
 - lib/swarm_sdk/context_compactor/metrics.rb
@@ -103,6 +104,9 @@ files:
 - lib/swarm_sdk/log_collector.rb
 - lib/swarm_sdk/log_stream.rb
 - lib/swarm_sdk/markdown_parser.rb
+- lib/swarm_sdk/model_aliases.json
+- lib/swarm_sdk/models.json
+- lib/swarm_sdk/models.rb
 - lib/swarm_sdk/node/agent_config.rb
 - lib/swarm_sdk/node/builder.rb
 - lib/swarm_sdk/node/transformer_executor.rb