llm_cost_tracker 0.15.2 → 0.15.3

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 380c06ddfdda1e5df94061a377dc703fd1f52c4b5ab92ab1f786ef1fce591be1
-  data.tar.gz: e58224c8eb4d538faea70fe867d33be5000611a35ec58ab33feb72d39fca9ef6
+  metadata.gz: 2fcb73e485dc1b20fed09f36b2095911d5268e16e633645329136acbe5ef3d3f
+  data.tar.gz: 4d3622d57bd89110efa3c824fcf87bc5b5a7f785ee371b7bc42a8351219f8df5
 SHA512:
-  metadata.gz: 2cd2b3c33e60e7638c58044220f2e43336be78f2cd085d253cdf99badb9af8586f411d1ced9e332a7255b81f08267407a52e7610b5dedc33fc2623cbca681a34
-  data.tar.gz: 70c17315f8a243cc59ebaf036f480151320a4ecb3adc53d152f5f15bfe50b51d81fdc0722e784d026d540e94d64d9a64443cc6f8cf51bcaf5585093b80ee049a
+  metadata.gz: 0ac48c68f138cfe1c7707b7ed1576c8c033a0ccbe0930f500bcc57d0c458c41656889095ec20cfc66b9913c1468fedfae37fac959f5696de678779d6e09fe1cd
+  data.tar.gz: e9fc0c48df5d63382e2c98f150ae2ed25130004a80514d540a1e83a06b9b2cbbf838f4ca29ef931609d0d7cf3c95f24252b8be0bf50f03945a99bae95611d4e8

data/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/). Versioning: [SemVer](https://semver.org/spec/v2.0.0.html).
 
+## [0.15.3] - 2026-11-11
+
+### Changed
+
+- Bundled prices add Claude Fable 5.2 and Mythos 5.2, Gemini 4 Carbon, Gemini Omni 1.2 Flash, GPT-6-Cyber, GPT-6 Rosalind and GPT-Image-2.6, and follow Mistral's Large 4 price cut.
+
+### Fixed
+
+- openai gem 0.115+ `connection.response.create` on a Responses WebSocket keeps its tags and `:block_requests` budget check.
+- RubyLLM 2.2: an attempt the provider answered with a 5xx before any output is recorded again, with unknown cost.
+- RubyLLM 2.2: transcription without usage is priced by its duration again, and speech without usage is no longer recorded as free.
+- RubyLLM 2.2: Claude on Bedrock, now called through InvokeModel, keeps its inference profile, data residency and service tier.
+
 ## [0.15.2] - 2026-10-29
 
 ### Changed

data/lib/llm_cost_tracker/integrations/openai/patches.rb

@@ -71,6 +71,12 @@ module LlmCostTracker
         end
       end
 
+      module ResponsesConnectionCreatePatch
+        def send_response_create(event)
+          llm_cost_tracker_capture.sending(event) { super }
+        end
+      end
+
       module ChatCompletionsPatch
         include PatchBuilder.build(record_method: :record_response, methods: %i[create])
         include PatchBuilder.build_stream(methods: %i[stream stream_raw])

data/lib/llm_cost_tracker/integrations/openai.rb

@@ -31,6 +31,11 @@ module LlmCostTracker
                          with: ResponsesConnectionPatch,
                          optional: true,
                          since: "0.86.0"),
+            patch_target("OpenAI::Responses::Connection",
+                         with: ResponsesConnectionCreatePatch,
+                         optional: true,
+                         skip_when_methods_missing: true,
+                         since: "0.86.0"),
             patch_target("OpenAI::Resources::Chat::Completions", with: ChatCompletionsPatch),
             patch_target("OpenAI::Resources::Completions", with: CreatePatch, optional: true),
             patch_target("OpenAI::Resources::Embeddings", with: CreatePatch, optional: true),

data/lib/llm_cost_tracker/integrations/ruby_llm/v2/attempt/token_counts.rb

@@ -7,6 +7,7 @@ module LlmCostTracker
         class Attempt
           class TokenCounts
             REFUSED = { input_tokens: 0, output_tokens: 0 }.freeze
+            NON_TOKEN_COUNTS = %i[characters audio_seconds search_units].freeze
 
             def initialize(usage, payload, raw:, request:)
               @usage = usage
@@ -62,7 +63,7 @@ module LlmCostTracker
 
             private
 
-            def known? = @tokens.to_h.any? || !@tokens.reported_cost.nil?
+            def known? = @tokens.to_h.except(*NON_TOKEN_COUNTS).any? || !@tokens.reported_cost.nil?
 
             def usage_source(known, result)
               return Usage::Source::SDK_RESPONSE if known

data/lib/llm_cost_tracker/integrations/ruby_llm/v2/attempt.rb

@@ -8,6 +8,8 @@ module LlmCostTracker
     module RubyLlm
       module V2
         class Attempt
+          BEDROCK_CLAUDE = /(?:\A|\.)anthropic\./
+
           attr_reader :usage, :latency_ms
 
           class << self
@@ -98,9 +100,9 @@ module LlmCostTracker
             return unless @events
 
             context = raw_context(raw, request).merge(response_status: 200, events: @events)
-            Attempt.stream_parsers.lazy.filter_map { |parser| parser.parse_stream(**context) }
-                   .find { |parsed| parsed.usage_source != Usage::Source::UNKNOWN }
-                   &.with(provider: usage[:provider], usage_source: Usage::Source::SDK_RESPONSE)
+            event = Attempt.stream_parsers.lazy.filter_map { |parser| parser.parse_stream(**context) }
+                           .find { |parsed| parsed.usage_source != Usage::Source::UNKNOWN }
+            sdk_event(event, usage, raw)
           end
 
           def parsed_event(usage, own, request)
@@ -108,8 +110,16 @@ module LlmCostTracker
             return unless body.is_a?(Hash)
 
             response = raw_context(own, request).merge(response_status: own.status, response_body: body)
-            event = body_event(usage, own, request, response)
-            event&.with(provider: usage[:provider], usage_source: Usage::Source::SDK_RESPONSE)
+            sdk_event(body_event(usage, own, request, response), usage, own)
+          end
+
+          def sdk_event(event, usage, raw)
+            event = event&.with(provider: usage[:provider], usage_source: Usage::Source::SDK_RESPONSE)
+            return event unless event && usage[:provider] == "bedrock" && usage[:model].to_s.match?(BEDROCK_CLAUDE)
+
+            request = { model: usage[:model], service_tier: raw.try(:headers).try(:[], "x-amzn-bedrock-service-tier") }
+            event.with(model: usage[:model],
+                       pricing_mode: Providers::Anthropic::UsageExtractor.pricing_mode(request: request, usage: nil))
           end
 
           def body_event(usage, own, request, response)

data/lib/llm_cost_tracker/integrations/ruby_llm/v2/frame.rb

@@ -36,6 +36,8 @@ module LlmCostTracker
           end
 
           def observe(seam, value, provider)
+            return @error_status = value.try(:status) if seam == :parse_error
+
             @provider = provider
             case seam
             when :parse_completion_body then @raw = value unless value.is_a?(Hash)
@@ -55,6 +57,8 @@ module LlmCostTracker
           end
 
           def add_attempt(usage)
+            usage = usage.merge(tokens: RubyLLM::Tokens.new) if maybe_billed?(usage)
+            @error_status = nil
             raw = @raw unless usage[:status] == :succeeded && usage[:tokens].to_h.empty?
             @reported << [usage, { latency_ms: @latency_ms, events: @window&.events, raw: raw }]
             @window = nil
@@ -68,6 +72,13 @@ module LlmCostTracker
               Attempt.new(usage, payload, final: index == final, response: @response, provider: @provider, **reported)
             end
           end
+
+          private
+
+          def maybe_billed?(usage)
+            usage[:status] != :succeeded && (500..599).cover?(@error_status.to_i) &&
+              usage[:tokens].to_h == Attempt::TokenCounts::REFUSED
+          end
         end
       end
     end

data/lib/llm_cost_tracker/integrations/ruby_llm/v2/seams.rb

@@ -40,6 +40,12 @@ module LlmCostTracker
         BatchBridge = Module.new do
           %i[messages results].each { |name| define_method(name) { V2.collect(self) { super() } } }
         end
+        ErrorBridge = Module.new do
+          def parse_error(*, **options, &)
+            V2.observe(:parse_error, options[:response])
+            super
+          end
+        end
         BRIDGES = (SEAMS.keys - %i[stream_transcription transcribe messages results current_workflow]).to_h do |seam|
           bridge = Module.new do
             define_method(seam) do |value, *args, **options, &block|
@@ -63,16 +69,26 @@ module LlmCostTracker
                   owner.prepend(bridge) unless owner == bridge
                 end
               end
+              middleware = error_middleware
+              return unless middleware
+
+              middleware.singleton_class.prepend(ErrorBridge)
             end
 
             def missing
-              SEAMS.flat_map do |seam, targets|
+              seams = SEAMS.flat_map do |seam, targets|
                 targets.reject { |target| owner(target, seam) }.map { |target| "RubyLLM::#{target}##{seam}" }
               end
+              error_middleware ? seams : [*seams, "RubyLLM::Transport::ErrorMiddleware.parse_error"]
             end
 
             private
 
+            def error_middleware
+              middleware = "RubyLLM::Transport::ErrorMiddleware".safe_constantize
+              middleware if middleware.respond_to?(:parse_error)
+            end
+
             def owner(target, seam)
               "RubyLLM::#{target}".safe_constantize&.instance_method(seam)&.owner
             rescue NameError