dify_llm 1.9.1 → 1.14.1

data/lib/ruby_llm/providers/openai/capabilities.rb

@@ -3,13 +3,11 @@
 module RubyLLM
   module Providers
     class OpenAI
-      # Determines capabilities and pricing for OpenAI models
+      # Provider-level capability checks and narrow registry fallbacks.
       module Capabilities
         module_function
 
         MODEL_PATTERNS = {
-          dall_e: /^dall-e/,
-          chatgpt4o: /^chatgpt-4o/,
           gpt41: /^gpt-4\.1(?!-(?:mini|nano))/,
           gpt41_mini: /^gpt-4\.1-mini/,
           gpt41_nano: /^gpt-4\.1-nano/,
@@ -26,9 +24,9 @@ module RubyLLM
           gpt4o_realtime: /^gpt-4o-realtime/,
           gpt4o_search: /^gpt-4o-search/,
           gpt4o_transcribe: /^gpt-4o-transcribe/,
-          gpt5: /^gpt-5/,
-          gpt5_mini: /^gpt-5-mini/,
-          gpt5_nano: /^gpt-5-nano/,
+          gpt5: /^gpt-5(?!.*(?:mini|nano))/,
+          gpt5_mini: /^gpt-5.*mini/,
+          gpt5_nano: /^gpt-5.*nano/,
           o1: /^o1(?!-(?:mini|pro))/,
           o1_mini: /^o1-mini/,
           o1_pro: /^o1-pro/,
@@ -44,71 +42,6 @@ module RubyLLM
           moderation: /^(?:omni|text)-moderation/
         }.freeze
 
-        def context_window_for(model_id)
-          case model_family(model_id)
-          when 'gpt41', 'gpt41_mini', 'gpt41_nano' then 1_047_576
-          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'chatgpt4o', 'gpt4_turbo', 'gpt4o', 'gpt4o_audio', 'gpt4o_mini',
-               'gpt4o_mini_audio', 'gpt4o_mini_realtime', 'gpt4o_realtime',
-               'gpt4o_search', 'gpt4o_transcribe', 'gpt4o_mini_search', 'o1_mini' then 128_000
-          when 'gpt4' then 8_192
-          when 'gpt4o_mini_transcribe' then 16_000
-          when 'o1', 'o1_pro', 'o3_mini' then 200_000
-          when 'gpt35_turbo' then 16_385
-          when 'gpt4o_mini_tts', 'tts1', 'tts1_hd', 'whisper', 'moderation',
-               'embedding3_large', 'embedding3_small', 'embedding_ada' then nil
-          else 4_096
-          end
-        end
-
-        def max_tokens_for(model_id)
-          case model_family(model_id)
-          when 'gpt5', 'gpt5_mini', 'gpt5_nano' then 400_000
-          when 'gpt41', 'gpt41_mini', 'gpt41_nano' then 32_768
-          when 'chatgpt4o', 'gpt4o', 'gpt4o_mini', 'gpt4o_mini_search' then 16_384
-          when 'babbage', 'davinci' then 16_384 # rubocop:disable Lint/DuplicateBranch
-          when 'gpt4' then 8_192
-          when 'gpt35_turbo' then 4_096
-          when 'gpt4_turbo', 'gpt4o_realtime', 'gpt4o_mini_realtime' then 4_096 # rubocop:disable Lint/DuplicateBranch
-          when 'gpt4o_mini_transcribe' then 2_000
-          when 'o1', 'o1_pro', 'o3_mini' then 100_000
-          when 'o1_mini' then 65_536
-          when 'gpt4o_mini_tts', 'tts1', 'tts1_hd', 'whisper', 'moderation',
-               'embedding3_large', 'embedding3_small', 'embedding_ada' then nil
-          else 16_384 # rubocop:disable Lint/DuplicateBranch
-          end
-        end
-
-        def supports_vision?(model_id)
-          case model_family(model_id)
-          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'chatgpt4o', 'gpt4',
-               'gpt4_turbo', 'gpt4o', 'gpt4o_mini', 'o1', 'o1_pro', 'moderation', 'gpt4o_search',
-               'gpt4o_mini_search' then true
-          else false
-          end
-        end
-
-        def supports_functions?(model_id)
-          case model_family(model_id)
-          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'gpt4', 'gpt4_turbo', 'gpt4o',
-               'gpt4o_mini', 'o1', 'o1_pro', 'o3_mini' then true
-          when 'chatgpt4o', 'gpt35_turbo', 'o1_mini', 'gpt4o_mini_tts',
-               'gpt4o_transcribe', 'gpt4o_search', 'gpt4o_mini_search' then false
-          else false # rubocop:disable Lint/DuplicateBranch
-          end
-        end
-
-        def supports_structured_output?(model_id)
-          case model_family(model_id)
-          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'chatgpt4o', 'gpt4o',
-               'gpt4o_mini', 'o1', 'o1_pro', 'o3_mini' then true
-          else false
-          end
-        end
-
-        def supports_json_mode?(model_id)
-          supports_structured_output?(model_id)
-        end
-
         PRICES = {
           gpt5: { input: 1.25, output: 10.0, cached_input: 0.125 },
           gpt5_mini: { input: 0.25, output: 2.0, cached_input: 0.025 },
@@ -116,21 +49,19 @@ module RubyLLM
           gpt41: { input: 2.0, output: 8.0, cached_input: 0.5 },
           gpt41_mini: { input: 0.4, output: 1.6, cached_input: 0.1 },
           gpt41_nano: { input: 0.1, output: 0.4 },
-          chatgpt4o: { input: 5.0, output: 15.0 },
           gpt4: { input: 10.0, output: 30.0 },
           gpt4_turbo: { input: 10.0, output: 30.0 },
-          gpt45: { input: 75.0, output: 150.0 },
           gpt35_turbo: { input: 0.5, output: 1.5 },
           gpt4o: { input: 2.5, output: 10.0 },
-          gpt4o_audio: { input: 2.5, output: 10.0, audio_input: 40.0, audio_output: 80.0 },
+          gpt4o_audio: { input: 2.5, output: 10.0 },
           gpt4o_mini: { input: 0.15, output: 0.6 },
-          gpt4o_mini_audio: { input: 0.15, output: 0.6, audio_input: 10.0, audio_output: 20.0 },
+          gpt4o_mini_audio: { input: 0.15, output: 0.6 },
           gpt4o_mini_realtime: { input: 0.6, output: 2.4 },
-          gpt4o_mini_transcribe: { input: 1.25, output: 5.0, audio_input: 3.0 },
+          gpt4o_mini_transcribe: { input: 1.25, output: 5.0 },
           gpt4o_mini_tts: { input: 0.6, output: 12.0 },
           gpt4o_realtime: { input: 5.0, output: 20.0 },
           gpt4o_search: { input: 2.5, output: 10.0 },
-          gpt4o_transcribe: { input: 2.5, output: 10.0, audio_input: 6.0 },
+          gpt4o_transcribe: { input: 2.5, output: 10.0 },
           o1: { input: 15.0, output: 60.0 },
           o1_mini: { input: 1.1, output: 4.4 },
           o1_pro: { input: 150.0, output: 600.0 },
@@ -146,153 +77,126 @@ module RubyLLM
           moderation: { price: 0.0 }
         }.freeze
 
-        def model_family(model_id)
-          MODEL_PATTERNS.each do |family, pattern|
-            return family.to_s if model_id.match?(pattern)
-          end
-          'other'
+        def supports_tool_choice?(_model_id)
+          true
         end
 
-        def input_price_for(model_id)
-          family = model_family(model_id).to_sym
-          prices = PRICES.fetch(family, { input: default_input_price })
-          prices[:input] || prices[:price] || default_input_price
-        end
-
-        def cached_input_price_for(model_id)
-          family = model_family(model_id).to_sym
-          prices = PRICES.fetch(family, {})
-          prices[:cached_input]
-        end
-
-        def output_price_for(model_id)
-          family = model_family(model_id).to_sym
-          prices = PRICES.fetch(family, { output: default_output_price })
-          prices[:output] || prices[:price] || default_output_price
+        def supports_tool_parallel_control?(_model_id)
+          true
         end
 
-        def model_type(model_id)
+        def context_window_for(model_id)
           case model_family(model_id)
-          when /embedding/ then 'embedding'
-          when /^tts|whisper|gpt4o_(?:mini_)?(?:transcribe|tts)$/ then 'audio'
-          when 'moderation' then 'moderation'
-          when /dall/ then 'image'
-          else 'chat'
+          when 'gpt41', 'gpt41_mini', 'gpt41_nano' then 1_047_576
+          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt4_turbo', 'gpt4o', 'gpt4o_audio', 'gpt4o_mini',
+               'gpt4o_mini_audio', 'gpt4o_mini_realtime', 'gpt4o_realtime', 'gpt4o_search',
+               'gpt4o_transcribe', 'o1_mini' then 128_000
+          when 'gpt4' then 8_192
+          when 'gpt4o_mini_transcribe' then 16_000
+          when 'o1', 'o1_pro', 'o3_mini' then 200_000
+          when 'gpt35_turbo' then 16_385
+          when 'gpt4o_mini_tts', 'tts1', 'tts1_hd', 'whisper', 'moderation',
+               'embedding3_large', 'embedding3_small', 'embedding_ada' then nil
+          else 4_096
           end
         end
 
-        def default_input_price
-          0.50
+        def max_tokens_for(model_id)
+          case model_family(model_id)
+          when 'gpt5', 'gpt5_mini', 'gpt5_nano' then 400_000
+          when 'gpt41', 'gpt41_mini', 'gpt41_nano' then 32_768
+          when 'gpt4' then 8_192
+          when 'gpt35_turbo' then 4_096
+          when 'gpt4o_mini_transcribe' then 2_000
+          when 'o1', 'o1_pro', 'o3_mini' then 100_000
+          when 'o1_mini' then 65_536
+          when 'gpt4o_mini_tts', 'tts1', 'tts1_hd', 'whisper', 'moderation',
+               'embedding3_large', 'embedding3_small', 'embedding_ada' then nil
+          else 16_384
+          end
         end
 
-        def default_output_price
-          1.50
+        def critical_capabilities_for(model_id)
+          capabilities = []
+          capabilities << 'function_calling' if supports_functions?(model_id)
+          capabilities << 'structured_output' if supports_structured_output?(model_id)
+          capabilities << 'vision' if supports_vision?(model_id)
+          capabilities << 'reasoning' if model_id.match?(/o\d|gpt-5|codex/)
+          capabilities
         end
 
-        def format_display_name(model_id)
-          model_id.then { |id| humanize(id) }
-                  .then { |name| apply_special_formatting(name) }
-        end
+        def pricing_for(model_id)
+          standard_pricing = {
+            input_per_million: input_price_for(model_id),
+            output_per_million: output_price_for(model_id)
+          }
 
-        def humanize(id)
-          id.tr('-', ' ')
-            .split
-            .map(&:capitalize)
-            .join(' ')
-        end
+          cached_price = cached_input_price_for(model_id)
+          standard_pricing[:cached_input_per_million] = cached_price if cached_price
 
-        def apply_special_formatting(name)
-          name
-            .gsub(/(\d{4}) (\d{2}) (\d{2})/, '\1\2\3')
-            .gsub(/^(?:Gpt|Chatgpt|Tts|Dall E) /) { |m| special_prefix_format(m.strip) }
-            .gsub(/^O([13]) /, 'O\1-')
-            .gsub(/^O[13] Mini/, '\0'.tr(' ', '-'))
-            .gsub(/\d\.\d /, '\0'.sub(' ', '-'))
-            .gsub(/4o (?=Mini|Preview|Turbo|Audio|Realtime|Transcribe|Tts)/, '4o-')
-            .gsub(/\bHd\b/, 'HD')
-            .gsub(/(?:Omni|Text) Moderation/, '\0'.tr(' ', '-'))
-            .gsub('Text Embedding', 'text-embedding-')
+          { text_tokens: { standard: standard_pricing } }
         end
 
-        def special_prefix_format(prefix)
-          case prefix # rubocop:disable Style/HashLikeCase
-          when 'Gpt' then 'GPT-'
-          when 'Chatgpt' then 'ChatGPT-'
-          when 'Tts' then 'TTS-'
-          when 'Dall E' then 'DALL-E-'
+        def model_family(model_id)
+          MODEL_PATTERNS.each do |family, pattern|
+            return family.to_s if model_id.match?(pattern)
           end
+
+          'other'
         end
 
-        def self.normalize_temperature(temperature, model_id)
-          if model_id.match?(/^(o\d|gpt-5)/)
-            RubyLLM.logger.debug "Model #{model_id} requires temperature=1.0, ignoring provided value"
-            1.0
-          elsif model_id.match?(/-search/)
-            RubyLLM.logger.debug "Model #{model_id} does not accept temperature parameter, removing"
-            nil
+        def supports_vision?(model_id)
+          case model_family(model_id)
+          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'gpt4',
+               'gpt4_turbo', 'gpt4o', 'gpt4o_mini', 'o1', 'o1_pro', 'moderation', 'gpt4o_search'
+            true
           else
-            temperature
+            false
           end
         end
 
-        def modalities_for(model_id)
-          modalities = {
-            input: ['text'],
-            output: ['text']
-          }
-
-          # Vision support
-          modalities[:input] << 'image' if supports_vision?(model_id)
-          modalities[:input] << 'audio' if model_id.match?(/whisper|audio|tts|transcribe/)
-          modalities[:input] << 'pdf' if supports_vision?(model_id)
-          modalities[:output] << 'audio' if model_id.match?(/tts|audio/)
-          modalities[:output] << 'image' if model_id.match?(/dall-e|image/)
-          modalities[:output] << 'embeddings' if model_id.match?(/embedding/)
-          modalities[:output] << 'moderation' if model_id.match?(/moderation/)
-
-          modalities
+        def supports_functions?(model_id)
+          case model_family(model_id)
+          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'gpt4',
+               'gpt4_turbo', 'gpt4o', 'gpt4o_mini', 'o1', 'o1_pro', 'o3_mini'
+            true
+          else
+            false
+          end
         end
 
-        def capabilities_for(model_id) # rubocop:disable Metrics/PerceivedComplexity
-          capabilities = []
-
-          capabilities << 'streaming' unless model_id.match?(/moderation|embedding/)
-          capabilities << 'function_calling' if supports_functions?(model_id)
-          capabilities << 'structured_output' if supports_json_mode?(model_id)
-          capabilities << 'batch' if model_id.match?(/embedding|batch/)
-          capabilities << 'reasoning' if model_id.match?(/o\d|gpt-5|codex/)
-
-          if model_id.match?(/gpt-4-turbo|gpt-4o/)
-            capabilities << 'image_generation' if model_id.match?(/vision/)
-            capabilities << 'speech_generation' if model_id.match?(/audio/)
-            capabilities << 'transcription' if model_id.match?(/audio/)
+        def supports_structured_output?(model_id)
+          case model_family(model_id)
+          when 'gpt5', 'gpt5_mini', 'gpt5_nano', 'gpt41', 'gpt41_mini', 'gpt41_nano', 'gpt4o',
+               'gpt4o_mini', 'o1', 'o1_pro', 'o3_mini'
+            true
+          else
+            false
           end
-
-          capabilities
         end
 
-        def pricing_for(model_id)
-          standard_pricing = {
-            input_per_million: input_price_for(model_id),
-            output_per_million: output_price_for(model_id)
-          }
-
-          if respond_to?(:cached_input_price_for)
-            cached_price = cached_input_price_for(model_id)
-            standard_pricing[:cached_input_per_million] = cached_price if cached_price
-          end
+        def input_price_for(model_id)
+          price_for(model_id, :input, 0.50)
+        end
 
-          pricing = { text_tokens: { standard: standard_pricing } }
+        def output_price_for(model_id)
+          price_for(model_id, :output, 1.50)
+        end
 
-          if model_id.match?(/embedding|batch/)
-            pricing[:text_tokens][:batch] = {
-              input_per_million: standard_pricing[:input_per_million] * 0.5,
-              output_per_million: standard_pricing[:output_per_million] * 0.5
-            }
-          end
+        def cached_input_price_for(model_id)
+          family = model_family(model_id).to_sym
+          PRICES.fetch(family, {})[:cached_input]
+        end
 
-          pricing
+        def price_for(model_id, key, fallback)
+          family = model_family(model_id).to_sym
+          prices = PRICES.fetch(family, { key => fallback })
+          prices[key] || prices[:price] || fallback
         end
+
+        module_function :context_window_for, :max_tokens_for, :critical_capabilities_for, :pricing_for,
+                        :model_family, :supports_vision?, :supports_functions?, :supports_structured_output?,
+                        :input_price_for, :output_price_for, :cached_input_price_for, :price_for
       end
     end
   end

data/lib/ruby_llm/providers/openai/chat.rb

@@ -11,7 +11,10 @@ module RubyLLM
 
         module_function
 
-        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil) # rubocop:disable Metrics/ParameterLists
+        # rubocop:disable Metrics/ParameterLists,Metrics/PerceivedComplexity
+        def render_payload(messages, tools:, temperature:, model:, stream: false, schema: nil,
+                           thinking: nil, tool_prefs: nil)
+          tool_prefs ||= {}
           payload = {
             model: model.id,
             messages: format_messages(messages),
@@ -19,24 +22,34 @@ module RubyLLM
           }
 
           payload[:temperature] = temperature unless temperature.nil?
-          payload[:tools] = tools.map { |_, tool| tool_for(tool) } if tools.any?
+          if tools.any?
+            payload[:tools] = tools.map { |_, tool| tool_for(tool) }
+            payload[:tool_choice] = build_tool_choice(tool_prefs[:choice]) unless tool_prefs[:choice].nil?
+            payload[:parallel_tool_calls] = tool_prefs[:calls] == :many unless tool_prefs[:calls].nil?
+          end
 
           if schema
-            strict = schema[:strict] != false
+            schema_name = schema[:name]
+            schema_def = schema[:schema]
+            strict = schema[:strict]
 
             payload[:response_format] = {
               type: 'json_schema',
               json_schema: {
-                name: 'response',
-                schema: schema,
+                name: schema_name,
+                schema: schema_def,
                 strict: strict
               }
             }
           end
 
+          effort = resolve_effort(thinking)
+          payload[:reasoning_effort] = effort if effort
+
           payload[:stream_options] = { include_usage: true } if stream
           payload
         end
+        # rubocop:enable Metrics/ParameterLists,Metrics/PerceivedComplexity
 
         def parse_completion_response(response)
           data = response.body
@@ -49,15 +62,21 @@ module RubyLLM
 
           usage = data['usage'] || {}
           cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          thinking_tokens = usage.dig('completion_tokens_details', 'reasoning_tokens')
+          content, thinking_from_blocks = extract_content_and_thinking(message_data['content'])
+          thinking_text = thinking_from_blocks || extract_thinking_text(message_data)
+          thinking_signature = extract_thinking_signature(message_data)
 
           Message.new(
             role: :assistant,
-            content: message_data['content'],
+            content: content,
+            thinking: Thinking.build(text: thinking_text, signature: thinking_signature),
             tool_calls: parse_tool_calls(message_data['tool_calls']),
             input_tokens: usage['prompt_tokens'],
             output_tokens: usage['completion_tokens'],
             cached_tokens: cached_tokens,
             cache_creation_tokens: 0,
+            thinking_tokens: thinking_tokens,
             model_id: data['model'],
             raw: response
           )
@@ -70,7 +89,7 @@ module RubyLLM
               content: Media.format_content(msg.content),
               tool_calls: format_tool_calls(msg.tool_calls),
               tool_call_id: msg.tool_call_id
-            }.compact
+            }.compact.merge(format_thinking(msg))
           end
         end
 
@@ -82,6 +101,81 @@ module RubyLLM
             role.to_s
           end
         end
+
+        def resolve_effort(thinking)
+          return nil unless thinking
+
+          thinking.respond_to?(:effort) ? thinking.effort : thinking
+        end
+
+        def format_thinking(msg)
+          return {} unless msg.role == :assistant
+
+          thinking = msg.thinking
+          return {} unless thinking
+
+          payload = {}
+          if thinking.text
+            payload[:reasoning] = thinking.text
+            payload[:reasoning_content] = thinking.text
+          end
+          payload[:reasoning_signature] = thinking.signature if thinking.signature
+          payload
+        end
+
+        def extract_thinking_text(message_data)
+          candidate = message_data['reasoning_content'] || message_data['reasoning'] || message_data['thinking']
+          candidate.is_a?(String) ? candidate : nil
+        end
+
+        def extract_thinking_signature(message_data)
+          candidate = message_data['reasoning_signature'] || message_data['signature']
+          candidate.is_a?(String) ? candidate : nil
+        end
+
+        def extract_content_and_thinking(content)
+          return extract_think_tag_content(content) if content.is_a?(String)
+          return [content, nil] unless content.is_a?(Array)
+
+          text = extract_text_from_blocks(content)
+          thinking = extract_thinking_from_blocks(content)
+
+          [text.empty? ? nil : text, thinking.empty? ? nil : thinking]
+        end
+
+        def extract_text_from_blocks(blocks)
+          blocks.filter_map do |block|
+            block['text'] if block['type'] == 'text' && block['text'].is_a?(String)
+          end.join
+        end
+
+        def extract_thinking_from_blocks(blocks)
+          blocks.filter_map do |block|
+            next unless block['type'] == 'thinking'
+
+            extract_thinking_text_from_block(block)
+          end.join
+        end
+
+        def extract_thinking_text_from_block(block)
+          thinking_block = block['thinking']
+          return thinking_block if thinking_block.is_a?(String)
+
+          if thinking_block.is_a?(Array)
+            return thinking_block.filter_map { |item| item['text'] if item['type'] == 'text' }.join
+          end
+
+          block['text'] if block['text'].is_a?(String)
+        end
+
+        def extract_think_tag_content(text)
+          return [text, nil] unless text.include?('<think>')
+
+          thinking = text.scan(%r{<think>(.*?)</think>}m).join
+          content = text.gsub(%r{<think>.*?</think>}m, '').strip
+
+          [content.empty? ? nil : content, thinking.empty? ? nil : thinking]
+        end
       end
     end
   end

data/lib/ruby_llm/providers/openai/media.rb

@@ -8,7 +8,10 @@ module RubyLLM
         module_function
 
         def format_content(content) # rubocop:disable Metrics/PerceivedComplexity
-          return content.value if content.is_a?(RubyLLM::Content::Raw)
+          if content.is_a?(RubyLLM::Content::Raw)
+            value = content.value
+            return value.is_a?(Hash) ? value.to_json : value
+          end
           return content.to_json if content.is_a?(Hash) || content.is_a?(Array)
           return content unless content.is_a?(Content)
 
@@ -37,7 +40,7 @@ module RubyLLM
           {
             type: 'image_url',
             image_url: {
-              url: image.url? ? image.source : image.for_llm
+              url: image.url? ? image.source.to_s : image.for_llm
             }
           }
         end

data/lib/ruby_llm/providers/openai/models.rb

@@ -17,14 +17,12 @@ module RubyLLM
 
             Model::Info.new(
               id: model_id,
-              name: capabilities.format_display_name(model_id),
+              name: model_id,
               provider: slug,
-              family: capabilities.model_family(model_id),
               created_at: model_data['created'] ? Time.at(model_data['created']) : nil,
               context_window: capabilities.context_window_for(model_id),
               max_output_tokens: capabilities.max_tokens_for(model_id),
-              modalities: capabilities.modalities_for(model_id),
-              capabilities: capabilities.capabilities_for(model_id),
+              capabilities: capabilities.critical_capabilities_for(model_id),
               pricing: capabilities.pricing_for(model_id),
               metadata: {
                 object: model_data['object'],

data/lib/ruby_llm/providers/openai/streaming.rb

@@ -14,16 +14,24 @@ module RubyLLM
         def build_chunk(data)
           usage = data['usage'] || {}
           cached_tokens = usage.dig('prompt_tokens_details', 'cached_tokens')
+          delta = data.dig('choices', 0, 'delta') || {}
+          content_source = delta['content'] || data.dig('choices', 0, 'message', 'content')
+          content, thinking_from_blocks = OpenAI::Chat.extract_content_and_thinking(content_source)
 
           Chunk.new(
             role: :assistant,
             model_id: data['model'],
-            content: data.dig('choices', 0, 'delta', 'content'),
-            tool_calls: parse_tool_calls(data.dig('choices', 0, 'delta', 'tool_calls'), parse_arguments: false),
+            content: content,
+            thinking: Thinking.build(
+              text: thinking_from_blocks || delta['reasoning_content'] || delta['reasoning'],
+              signature: delta['reasoning_signature']
+            ),
+            tool_calls: parse_tool_calls(delta['tool_calls'], parse_arguments: false),
             input_tokens: usage['prompt_tokens'],
             output_tokens: usage['completion_tokens'],
             cached_tokens: cached_tokens,
-            cache_creation_tokens: 0
+            cache_creation_tokens: 0,
+            thinking_tokens: usage.dig('completion_tokens_details', 'reasoning_tokens')
           )
         end
 

data/lib/ruby_llm/providers/openai/temperature.rb

@@ -0,0 +1,28 @@
+# frozen_string_literal: true
+
+module RubyLLM
+  module Providers
+    class OpenAI
+      # Normalizes temperature for OpenAI models with provider-specific requirements.
+      module Temperature
+        module_function
+
+        def normalize(temperature, model_id)
+          if model_id.match?(/^(o\d|gpt-5)/) && !temperature.nil? && !temperature_close_to_one?(temperature)
+            RubyLLM.logger.debug { "Model #{model_id} requires temperature=1.0, setting that instead." }
+            1.0
+          elsif model_id.include?('-search')
+            RubyLLM.logger.debug { "Model #{model_id} does not accept temperature parameter, removing" }
+            nil
+          else
+            temperature
+          end
+        end
+
+        def temperature_close_to_one?(temperature)
+          (temperature.to_f - 1.0).abs <= Float::EPSILON
+        end
+      end
+    end
+  end
+end