llm_cost_tracker 0.12.0 → 0.13.0

data/lib/llm_cost_tracker/prices.json

@@ -4,16 +4,16 @@
     "min_gem_version": "0.4.0",
     "schema_version": 1,
     "source_urls": [
-      "https://developers.openai.com/api/docs/pricing",
       "https://platform.claude.com/docs/en/about-claude/pricing",
       "https://ai.google.dev/gemini-api/docs/pricing",
-      "https://console.groq.com/docs/models",
+      "https://groq.com/pricing",
       "https://console.groq.com/docs/prompt-caching",
       "https://console.groq.com/docs/flex-processing",
-      "https://console.groq.com/docs/service-tiers"
+      "https://developers.openai.com/api/docs/pricing",
+      "https://openrouter.ai/api/v1/models"
     ],
     "unit": "1M tokens",
-    "updated_at": "2026-06-03"
+    "updated_at": "2026-06-26"
   },
   "models": {
     "anthropic/claude-haiku-3": {
@@ -34,15 +34,6 @@
       "input": 1.0,
       "output": 5.0
     },
-    "anthropic/claude-opus-4-1": {
-      "batch_input": 7.5,
-      "batch_output": 37.5,
-      "cache_read_input": 1.5,
-      "cache_write_extended_input": 30.0,
-      "cache_write_input": 18.75,
-      "input": 15.0,
-      "output": 75.0
-    },
     "anthropic/claude-opus-4-5": {
       "batch_input": 2.5,
       "batch_output": 12.5,
@@ -91,6 +82,16 @@
       "data_residency_cache_write_input": 6.875,
       "data_residency_input": 5.5,
       "data_residency_output": 27.5,
+      "fast_cache_read_input": 3.0,
+      "fast_cache_write_extended_input": 60.0,
+      "fast_cache_write_input": 37.5,
+      "fast_data_residency_cache_read_input": 3.3,
+      "fast_data_residency_cache_write_extended_input": 66.0,
+      "fast_data_residency_cache_write_input": 41.25,
+      "fast_data_residency_input": 33.0,
+      "fast_data_residency_output": 165.0,
+      "fast_input": 30.0,
+      "fast_output": 150.0,
       "input": 5.0,
       "output": 25.0
     },
@@ -107,6 +108,16 @@
       "data_residency_cache_write_input": 6.875,
       "data_residency_input": 5.5,
       "data_residency_output": 27.5,
+      "fast_cache_read_input": 1.0,
+      "fast_cache_write_extended_input": 20.0,
+      "fast_cache_write_input": 12.5,
+      "fast_data_residency_cache_read_input": 1.1,
+      "fast_data_residency_cache_write_extended_input": 22.0,
+      "fast_data_residency_cache_write_input": 13.75,
+      "fast_data_residency_input": 11.0,
+      "fast_data_residency_output": 55.0,
+      "fast_input": 10.0,
+      "fast_output": 50.0,
       "input": 5.0,
       "output": 25.0
     },
@@ -377,6 +388,14 @@
       "on_demand_output": 0.79,
       "output": 0.79
     },
+    "groq/meta-llama/llama-4-scout-17b-16e-instruct": {
+      "flex_input": 0.11,
+      "flex_output": 0.34,
+      "input": 0.11,
+      "on_demand_input": 0.11,
+      "on_demand_output": 0.34,
+      "output": 0.34
+    },
     "groq/openai/gpt-oss-120b": {
       "cache_read_input": 0.075,
       "flex_cache_read_input": 0.075,
@@ -399,6 +418,22 @@
       "on_demand_output": 0.3,
       "output": 0.3
     },
+    "groq/qwen/qwen3-32b": {
+      "flex_input": 0.29,
+      "flex_output": 0.59,
+      "input": 0.29,
+      "on_demand_input": 0.29,
+      "on_demand_output": 0.59,
+      "output": 0.59
+    },
+    "groq/qwen/qwen3.6-27b": {
+      "flex_input": 0.6,
+      "flex_output": 3.0,
+      "input": 0.6,
+      "on_demand_input": 0.6,
+      "on_demand_output": 3.0,
+      "output": 3.0
+    },
     "openai/chatgpt-4o-latest": {
       "input": 5.0,
       "output": 15.0
@@ -1083,6 +1118,12 @@
       "input": 0.8,
       "output": 4.0
     },
+    "openrouter/anthropic/claude-fable-5": {
+      "cache_read_input": 1.0,
+      "cache_write_input": 12.5,
+      "input": 10.0,
+      "output": 50.0
+    },
     "openrouter/anthropic/claude-haiku-4.5": {
       "cache_read_input": 0.1,
       "cache_write_input": 1.25,
@@ -1175,8 +1216,8 @@
     },
     "openrouter/arcee-ai/trinity-large-thinking": {
       "cache_read_input": 0.06,
-      "input": 0.22,
-      "output": 0.85
+      "input": 0.25,
+      "output": 0.8
     },
     "openrouter/arcee-ai/trinity-mini": {
       "input": 0.045,
@@ -1275,7 +1316,7 @@
       "output": 2.15
     },
     "openrouter/deepseek/deepseek-r1-distill-llama-70b": {
-      "input": 0.7,
+      "input": 0.8,
       "output": 0.8
     },
     "openrouter/deepseek/deepseek-r1-distill-qwen-32b": {
@@ -1288,6 +1329,7 @@
       "output": 0.95
     },
     "openrouter/deepseek/deepseek-v3.2": {
+      "cache_read_input": 0.02288,
       "input": 0.2288,
       "output": 0.3432
     },
@@ -1301,9 +1343,9 @@
       "output": 0.431
     },
     "openrouter/deepseek/deepseek-v4-flash": {
-      "cache_read_input": 0.0197,
-      "input": 0.0983,
-      "output": 0.1966
+      "cache_read_input": 0.02,
+      "input": 0.09,
+      "output": 0.18
     },
     "openrouter/deepseek/deepseek-v4-pro": {
       "cache_read_input": 0.003625,
@@ -1392,6 +1434,14 @@
       "input": 0.5,
       "output": 3.0
     },
+    "openrouter/google/gemini-3-pro-image": {
+      "audio_input": 2.0,
+      "cache_read_input": 0.2,
+      "cache_write_input": 0.375,
+      "image_input": 2.0,
+      "input": 2.0,
+      "output": 12.0
+    },
     "openrouter/google/gemini-3-pro-image-preview": {
       "audio_input": 2.0,
       "cache_read_input": 0.2,
@@ -1400,6 +1450,10 @@
       "input": 2.0,
       "output": 12.0
     },
+    "openrouter/google/gemini-3.1-flash-image": {
+      "input": 0.5,
+      "output": 3.0
+    },
     "openrouter/google/gemini-3.1-flash-image-preview": {
       "input": 0.5,
       "output": 3.0
@@ -1449,16 +1503,16 @@
       "output": 0.65
     },
     "openrouter/google/gemma-3-12b-it": {
-      "input": 0.04,
-      "output": 0.13
+      "input": 0.05,
+      "output": 0.15
     },
     "openrouter/google/gemma-3-27b-it": {
       "input": 0.08,
       "output": 0.16
     },
     "openrouter/google/gemma-3-4b-it": {
-      "input": 0.04,
-      "output": 0.08
+      "input": 0.05,
+      "output": 0.1
     },
     "openrouter/google/gemma-3n-e4b-it": {
       "input": 0.06,
@@ -1469,8 +1523,9 @@
       "output": 0.33
     },
     "openrouter/google/gemma-4-31b-it": {
+      "cache_read_input": 0.09,
       "input": 0.12,
-      "output": 0.37
+      "output": 0.35
     },
     "openrouter/gryphe/mythomax-l2-13b": {
       "input": 0.06,
@@ -1531,8 +1586,8 @@
       "output": 0.74
     },
     "openrouter/meta-llama/llama-3-8b-instruct": {
-      "input": 0.04,
-      "output": 0.04
+      "input": 0.14,
+      "output": 0.14
     },
     "openrouter/meta-llama/llama-3.1-70b-instruct": {
       "input": 0.4,
@@ -1540,11 +1595,11 @@
     },
     "openrouter/meta-llama/llama-3.1-8b-instruct": {
       "input": 0.02,
-      "output": 0.05
+      "output": 0.03
     },
     "openrouter/meta-llama/llama-3.2-11b-vision-instruct": {
-      "input": 0.245,
-      "output": 0.245
+      "input": 0.345,
+      "output": 0.345
     },
     "openrouter/meta-llama/llama-3.2-1b-instruct": {
       "input": 0.027,
@@ -1563,7 +1618,7 @@
       "output": 0.6
     },
     "openrouter/meta-llama/llama-4-scout": {
-      "input": 0.08,
+      "input": 0.1,
       "output": 0.3
     },
     "openrouter/meta-llama/llama-guard-3-8b": {
@@ -1575,7 +1630,7 @@
       "output": 0.18
     },
     "openrouter/microsoft/phi-4": {
-      "input": 0.065,
+      "input": 0.07,
       "output": 0.14
     },
     "openrouter/microsoft/phi-4-mini-instruct": {
@@ -1611,12 +1666,12 @@
       "output": 0.95
     },
     "openrouter/minimax/minimax-m2.5": {
-      "input": 0.15,
-      "output": 1.15
+      "input": 0.12,
+      "output": 0.48
     },
     "openrouter/minimax/minimax-m2.7": {
-      "input": 0.279,
-      "output": 1.2
+      "input": 0.18,
+      "output": 0.72
     },
     "openrouter/minimax/minimax-m3": {
       "cache_read_input": 0.06,
@@ -1747,18 +1802,23 @@
       "output": 2.5
     },
     "openrouter/moonshotai/kimi-k2-thinking": {
+      "cache_read_input": 0.6,
       "input": 0.6,
       "output": 2.5
     },
     "openrouter/moonshotai/kimi-k2.5": {
-      "cache_read_input": 0.09,
-      "input": 0.4,
-      "output": 1.9
+      "input": 0.375,
+      "output": 2.025
     },
     "openrouter/moonshotai/kimi-k2.6": {
       "cache_read_input": 0.144,
-      "input": 0.684,
-      "output": 3.42
+      "input": 0.66,
+      "output": 3.41
+    },
+    "openrouter/moonshotai/kimi-k2.7-code": {
+      "cache_read_input": 0.15,
+      "input": 0.74,
+      "output": 3.5
     },
     "openrouter/morph/morph-v3-fast": {
       "input": 0.8,
@@ -1772,6 +1832,11 @@
       "input": 0.135,
       "output": 0.5
     },
+    "openrouter/nex-agi/nex-n2-pro": {
+      "cache_read_input": 0.025,
+      "input": 0.25,
+      "output": 1.0
+    },
     "openrouter/nousresearch/hermes-2-pro-llama-3-8b": {
       "input": 0.14,
       "output": 0.14
@@ -1781,8 +1846,8 @@
       "output": 1.0
     },
     "openrouter/nousresearch/hermes-3-llama-3.1-70b": {
-      "input": 0.3,
-      "output": 0.3
+      "input": 0.7,
+      "output": 0.7
     },
     "openrouter/nousresearch/hermes-4-405b": {
       "input": 1.0,
@@ -1793,7 +1858,7 @@
       "output": 0.4
     },
     "openrouter/nvidia/llama-3.3-nemotron-super-49b-v1.5": {
-      "input": 0.1,
+      "input": 0.4,
       "output": 0.4
     },
     "openrouter/nvidia/nemotron-3-nano-30b-a3b": {
@@ -1801,8 +1866,13 @@
       "output": 0.2
     },
     "openrouter/nvidia/nemotron-3-super-120b-a12b": {
-      "input": 0.09,
-      "output": 0.45
+      "input": 0.085,
+      "output": 0.4
+    },
+    "openrouter/nvidia/nemotron-3-ultra-550b-a55b": {
+      "cache_read_input": 0.1,
+      "input": 0.5,
+      "output": 2.2
     },
     "openrouter/nvidia/nemotron-nano-9b-v2": {
       "input": 0.04,
@@ -2042,15 +2112,15 @@
       "output": 30.0
     },
     "openrouter/openai/gpt-oss-120b": {
-      "input": 0.039,
-      "output": 0.18
+      "input": 0.03,
+      "output": 0.15
     },
     "openrouter/openai/gpt-oss-20b": {
       "input": 0.029,
       "output": 0.14
     },
     "openrouter/openai/gpt-oss-safeguard-20b": {
-      "cache_read_input": 0.037,
+      "cache_read_input": 0.0375,
       "input": 0.075,
       "output": 0.3
     },
@@ -2126,6 +2196,16 @@
       "input": 2.0,
       "output": 8.0
     },
+    "openrouter/poolside/laguna-m.1": {
+      "cache_read_input": 0.1,
+      "input": 0.2,
+      "output": 0.4
+    },
+    "openrouter/poolside/laguna-xs.2": {
+      "cache_read_input": 0.05,
+      "input": 0.1,
+      "output": 0.2
+    },
     "openrouter/prime-intellect/intellect-3": {
       "input": 0.2,
       "output": 1.1
@@ -2158,8 +2238,9 @@
       "output": 0.78
     },
     "openrouter/qwen/qwen2.5-vl-72b-instruct": {
-      "input": 0.25,
-      "output": 0.75
+      "cache_read_input": 0.4,
+      "input": 0.8,
+      "output": 1.0
     },
     "openrouter/qwen/qwen3-14b": {
       "input": 0.1,
@@ -2170,7 +2251,7 @@
       "output": 1.82
     },
     "openrouter/qwen/qwen3-235b-a22b-2507": {
-      "input": 0.071,
+      "input": 0.09,
       "output": 0.1
     },
     "openrouter/qwen/qwen3-235b-a22b-thinking-2507": {
@@ -2179,12 +2260,12 @@
       "output": 0.1
     },
     "openrouter/qwen/qwen3-30b-a3b": {
-      "input": 0.09,
-      "output": 0.45
+      "input": 0.12,
+      "output": 0.5
     },
     "openrouter/qwen/qwen3-30b-a3b-instruct-2507": {
-      "input": 0.0428,
-      "output": 0.1716
+      "input": 0.04815,
+      "output": 0.19305
     },
     "openrouter/qwen/qwen3-30b-a3b-thinking-2507": {
       "cache_read_input": 0.08,
@@ -2286,11 +2367,11 @@
       "output": 1.0
     },
     "openrouter/qwen/qwen3.5-397b-a17b": {
-      "input": 0.39,
-      "output": 2.34
+      "input": 0.385,
+      "output": 2.45
     },
     "openrouter/qwen/qwen3.5-9b": {
-      "input": 0.04,
+      "input": 0.1,
       "output": 0.15
     },
     "openrouter/qwen/qwen3.5-flash-02-23": {
@@ -2307,8 +2388,8 @@
       "output": 1.8
     },
     "openrouter/qwen/qwen3.6-27b": {
-      "input": 0.29,
-      "output": 3.2
+      "input": 0.2885,
+      "output": 3.17
     },
     "openrouter/qwen/qwen3.6-35b-a3b": {
       "input": 0.14,
@@ -2335,6 +2416,12 @@
       "input": 1.25,
       "output": 3.75
     },
+    "openrouter/qwen/qwen3.7-plus": {
+      "cache_read_input": 0.064,
+      "cache_write_input": 0.4,
+      "input": 0.32,
+      "output": 1.28
+    },
     "openrouter/rekaai/reka-edge": {
       "input": 0.1,
       "output": 0.1
@@ -2351,6 +2438,11 @@
       "input": 1.0,
       "output": 3.0
     },
+    "openrouter/sakana/fugu-ultra": {
+      "cache_read_input": 0.5,
+      "input": 5.0,
+      "output": 30.0
+    },
     "openrouter/sao10k/l3-euryale-70b": {
       "input": 1.48,
       "output": 1.48
@@ -2400,8 +2492,8 @@
       "output": 0.5
     },
     "openrouter/thedrummer/rocinante-12b": {
-      "input": 0.17,
-      "output": 0.43
+      "input": 0.25,
+      "output": 0.5
     },
     "openrouter/thedrummer/skyfall-36b-v2": {
       "cache_read_input": 0.25,
@@ -2432,8 +2524,8 @@
     },
     "openrouter/x-ai/grok-4.20-multi-agent": {
       "cache_read_input": 0.2,
-      "input": 2.0,
-      "output": 6.0
+      "input": 1.25,
+      "output": 2.5
     },
     "openrouter/x-ai/grok-4.3": {
       "cache_read_input": 0.2,
@@ -2461,8 +2553,7 @@
       "output": 3.0
     },
     "openrouter/xiaomi/mimo-v2.5": {
-      "cache_read_input": 0.0028,
-      "input": 0.14,
+      "input": 0.105,
       "output": 0.28
     },
     "openrouter/xiaomi/mimo-v2.5-pro": {
@@ -2480,8 +2571,8 @@
       "output": 2.2
     },
     "openrouter/z-ai/glm-4.5-air": {
-      "cache_read_input": 0.06,
-      "input": 0.125,
+      "cache_read_input": 0.025,
+      "input": 0.13,
       "output": 0.85
     },
     "openrouter/z-ai/glm-4.5v": {
@@ -2495,7 +2586,7 @@
       "output": 1.74
     },
     "openrouter/z-ai/glm-4.6v": {
-      "cache_read_input": 0.05,
+      "cache_read_input": 0.055,
       "input": 0.3,
       "output": 0.9
     },
@@ -2524,11 +2615,22 @@
       "input": 0.98,
       "output": 3.08
     },
+    "openrouter/z-ai/glm-5.2": {
+      "cache_read_input": 0.18,
+      "input": 0.95,
+      "output": 3.0
+    },
     "openrouter/z-ai/glm-5v-turbo": {
       "cache_read_input": 0.24,
       "input": 1.2,
       "output": 4.0
     },
+    "openrouter/~anthropic/claude-fable-latest": {
+      "cache_read_input": 1.0,
+      "cache_write_input": 12.5,
+      "input": 10.0,
+      "output": 50.0
+    },
     "openrouter/~anthropic/claude-haiku-latest": {
       "cache_read_input": 0.1,
       "cache_write_input": 1.25,
@@ -2565,8 +2667,8 @@
     },
     "openrouter/~moonshotai/kimi-latest": {
       "cache_read_input": 0.144,
-      "input": 0.684,
-      "output": 3.42
+      "input": 0.66,
+      "output": 3.41
     },
     "openrouter/~openai/gpt-latest": {
       "cache_read_input": 0.5,

data/lib/llm_cost_tracker/pricing/calculation.rb

@@ -10,7 +10,8 @@ module LlmCostTracker
   module Pricing
     class Calculation
       RATE_DENOMINATOR_TOKENS = Pricing::RATE_BASIS_QUANTITIES.fetch("per_million_tokens")
-      private_constant :RATE_DENOMINATOR_TOKENS
+      SNAPSHOT_SCHEMA_VERSION = 1
+      private_constant :RATE_DENOMINATOR_TOKENS, :SNAPSHOT_SCHEMA_VERSION
 
       def self.for(provider:, model:, tokens:, pricing_mode:, line_items: [], usage_source: nil)
         new(provider: provider,
@@ -61,7 +62,12 @@ module LlmCostTracker
       def snapshot
         return @snapshot if defined?(@snapshot)
 
-        @snapshot = priceable? ? build_snapshot : nil
+        @snapshot =
+          if priceable?
+            build_snapshot
+          elsif kept_service_lines.any?
+            build_service_snapshot
+          end
       end
 
       def cost
@@ -131,7 +137,7 @@ module LlmCostTracker
 
       def build_snapshot
         {
-          "schema_version" => 1,
+          "schema_version" => SNAPSHOT_SCHEMA_VERSION,
           "source" => match.source.name,
           "source_key" => match.key,
           "source_version" => match.source.version,
@@ -141,6 +147,18 @@ module LlmCostTracker
         }
       end
 
+      def build_service_snapshot
+        primary = kept_service_lines.first
+        {
+          "schema_version" => SNAPSHOT_SCHEMA_VERSION,
+          "source" => primary.price_source,
+          "source_version" => primary.price_source_version,
+          "matched_by" => "service_charges",
+          "currency" => cost.currency,
+          "rates" => service_charge_rates
+        }
+      end
+
       def token_charge_rates
         priced_token_line_items.each_with_object({}) do |line_item, rates|
           next if line_item.price_key.nil? || line_item.rate_amount.nil?
@@ -150,8 +168,8 @@ module LlmCostTracker
       end
 
       def service_charge_rates
-        priced_line_items.each_with_object({}) do |line_item, rates|
-          next if line_item.token? || line_item.price_key.nil? || line_item.rate_amount.nil?
+        kept_service_lines.each_with_object({}) do |line_item, rates|
+          next if line_item.price_key.nil? || line_item.rate_amount.nil?
 
           rates[line_item.price_key] ||= rate_entry(line_item.rate_amount, line_item.rate_quantity)
         end
@@ -162,7 +180,7 @@ module LlmCostTracker
       end
 
       def price_token(line_item)
-        dimension = dimension_for(line_item)
+        dimension = line_item.dimension
         return line_item unless dimension
         return line_item.with(cost_status: Charges::CostStatus::UNKNOWN) unless priceable?
 
@@ -210,30 +228,31 @@ module LlmCostTracker
         )
       end
 
-      def dimension_for(line_item)
-        Usage::Catalog.all.find do |dimension|
-          dimension.kind == line_item.kind &&
-            dimension.direction == line_item.direction &&
-            dimension.modality == line_item.modality &&
-            dimension.cache_state == line_item.cache_state &&
-            dimension.unit == line_item.unit
+      def kept_service_lines
+        return @kept_service_lines if defined?(@kept_service_lines)
+
+        @kept_service_lines = begin
+          priced_services = priced_line_items.reject(&:token?).select(&:priced?)
+          if priced_services.empty?
+            []
+          else
+            base_currency = base_currency_for(token_cost, priced_services)
+            matching, mismatched = priced_services.partition { |line| line.currency.to_s == base_currency.to_s }
+            warn_currency_mismatch(mismatched, base_currency) if mismatched.any?
+            matching
+          end
         end
       end
 
       def combine_service_lines
         cost = token_cost
-        priced_services = priced_line_items.reject(&:token?).select(&:priced?)
-        return cost if priced_services.empty?
-
-        base_currency = base_currency_for(cost, priced_services)
-        matching, mismatched = priced_services.partition { |line| line.currency.to_s == base_currency.to_s }
-        warn_currency_mismatch(mismatched, base_currency) if mismatched.any?
+        return cost if kept_service_lines.empty?
 
-        service_total = matching.sum(BigDecimal("0")) { |line| line.cost_value.round(8) }
+        service_total = kept_service_lines.sum(BigDecimal("0")) { |line| line.cost_value.round(8) }
         Charges::Cost.new(
           components: cost ? cost.components : {}.freeze,
           total: (cost&.total || BigDecimal("0")) + service_total,
-          currency: (cost&.currency || base_currency).to_s
+          currency: (cost&.currency || kept_service_lines.first.currency).to_s
         )
       end