CostCompassAn AlmanacBeta

LLM pricing, by provider

What the model APIs CostCompass tracks charge per million tokens, side by side: input, output, and the cache rates that decide most of an agent's bill. These are the list prices CostCompass uses to estimate spend, and each one is compared with the provider's own pricing page once a week. The log at the bottom lists every price change that check found.

Prices per 1M tokens, in US dollars, as each provider lists them. Last change .

Anthropic (Claude)

ModelInputOutput
Claude Fable 5.110.0050.00
Claude Fable 5.1 · cache read0.25—
Claude Fable 5.1 · cache write12.50—
Claude Fable 5.1 · cache write 1h20.00—
Claude Fable 5.28.0040.00
Claude Fable 5.2 · cache read0.20—
Claude Fable 5.2 · cache write10.00—
Claude Fable 5.2 · cache write 1h16.00—
Claude Haiku 5.50.100.50
Claude Haiku 5.5 · cache read0.01—
Claude Haiku 5.5 · cache write0.125—
Claude Haiku 5.5 · cache write 1h0.20—
Claude Mythos 510.0050.00
Claude Mythos 5 · cache read1.00—
Claude Mythos 5 · cache write12.50—
Claude Mythos 5 · cache write 1h20.00—
Claude Mythos 5.110.0050.00
Claude Mythos 5.1 · cache read0.25—
Claude Mythos 5.1 · cache write12.50—
Claude Mythos 5.1 · cache write 1h20.00—
Claude Mythos 5.28.0040.00
Claude Mythos 5.2 · cache read0.20—
Claude Mythos 5.2 · cache write10.00—
Claude Mythos 5.2 · cache write 1h16.00—
Claude Opus 5.54.0020.00
Claude Opus 5.5 · cache read0.20—
Claude Opus 5.5 · cache write5.00—
Claude Opus 5.5 · cache write 1h8.00—
Claude Sonnet 5.52.0010.00
Claude Sonnet 5.5 · cache read0.10—
Claude Sonnet 5.5 · cache write2.50—
Claude Sonnet 5.5 · cache write 1h4.00—

DeepSeek

ModelInputOutput
DeepSeek Flash0.301.20
DeepSeek Flash · cache read0.006—
DeepSeek Flash · cache read · off peak0.003—
DeepSeek Flash · off peak0.150.60
DeepSeek V4 Pro1.323.96
DeepSeek V4 Pro · cache read0.044—
DeepSeek V4 Pro · cache read · off peak0.022—
DeepSeek V4 Pro · off peak0.661.98

Rates are time-of-day. Peak is 01:00⁠–⁠04:00 and 06:00⁠–⁠10:00 UTC, Monday to Friday; every other hour — and all weekend — is off-peak at half the peak rate. Input is the cache-miss rate. Search and Fetch are priced per successful request, at any hour.

Google

ModelInputOutput
Gemini 2.5 Flash0.302.50
Gemini 2.5 Flash · cached input0.03—
Gemini 2.5 Flash · thinking output—2.50
Gemini 2.5 Flash Image0.30—
Gemini 2.5 Flash Image · image output—30.00
Gemini 2.5 Flash Lite0.100.40
Gemini 2.5 Flash Lite · cached input0.01—
Gemini 2.5 Pro1.2510.00
Gemini 2.5 Pro · cached input0.125—
Gemini 2.5 Pro · thinking output—10.00
Gemini 3 Pro Image2.0012.00
Gemini 3 Pro Image · image output—120.00
Gemini 3.1 Flash Image0.503.00
Gemini 3.1 Flash Image · image output—60.00
Gemini 3.1 Flash Lite0.251.50
Gemini 3.1 Flash Lite · cached input0.025—
Gemini 3.1 Flash Lite · thinking output—1.50
Gemini 3.1 Flash Lite Image0.251.50
Gemini 3.1 Flash Lite Image · image output—30.00
Gemini 3.1 Pro Preview2.0012.00
Gemini 3.1 Pro Preview · cached input0.20—
Gemini 3.1 Pro Preview · thinking output—12.00
Gemini 3.5 Flash1.509.00
Gemini 3.5 Flash · cached input0.15—
Gemini 3.5 Flash · thinking output—9.00
Gemini 3.5 Flash Lite0.302.50
Gemini 3.5 Flash Lite · cached input0.03—
Gemini 3.5 Flash Lite · thinking output—2.50
Gemini 3.6 Flashuntil , then 1.50 / 7.500.753.75
Gemini 3.6 Flash · cached inputuntil , then 0.15 / —0.075—
Gemini 3.6 Flash · thinking outputuntil , then — / 7.50—3.75
Gemini 3.7 Flashuntil , then 1.50 / 7.500.753.75
Gemini 3.7 Flash · cached inputuntil , then 0.15 / —0.075—
Gemini 3.7 Flash · thinking outputuntil , then — / 7.50—3.75
Gemini 3.8 Flashuntil , then 1.50 / 7.500.753.75
Gemini 3.8 Flash · cached inputuntil , then 0.15 / —0.075—
Gemini 3.8 Flash · thinking outputuntil , then — / 7.50—3.75
Gemini 3.8 Flash Lite Ttsuntil , then 1.00 / —0.50—
Gemini 3.8 Flash Lite Tts · audio outputuntil , then — / 12.00—6.00
Gemini 3.8 Flash Ttsuntil , then 1.00 / —0.50—
Gemini 3.8 Flash Tts · audio outputuntil , then — / 18.00—9.00
Gemini 4 Argonuntil , then 4.00 / 20.002.0010.00
Gemini 4 Argon · cached inputuntil , then 0.20 / —0.10—
Gemini 4 Argon · thinking outputuntil , then — / 20.00—10.00

OpenAI

ModelInputOutput
chat-latest5.0030.00
chat-latest · cache read0.50—
GPT-4.12.008.00
GPT-4.1 · cache read0.50—
GPT-4.1 mini0.401.60
GPT-4.1 mini · cache read0.10—
GPT-4o2.5010.00
GPT-4o · cache read1.25—
GPT-4o mini0.150.60
GPT-4o mini · cache read0.075—
GPT-51.2510.00
GPT-5 · cache read0.125—
GPT-5 mini0.252.00
GPT-5 mini · cache read0.025—
GPT-5 nano0.050.40
GPT-5 nano · cache read0.005—
GPT-5 Pro15.00120.00
GPT-5 Search Api1.2510.00
GPT-5 Search Api · cache read0.125—
GPT-5.11.2510.00
GPT-5.1 · cache read0.125—
GPT-5.21.7514.00
GPT-5.2 · cache read0.175—
GPT-5.2 Pro21.00168.00
GPT-5.3 Codex1.7514.00
GPT-5.3 Codex · cache read0.175—
GPT-5.42.5015.00
GPT-5.4 · cache read0.25—
GPT-5.4 mini0.754.50
GPT-5.4 mini · cache read0.075—
GPT-5.4 nano0.201.25
GPT-5.4 nano · cache read0.02—
GPT-5.4 Pro30.00180.00
GPT-5.55.0030.00
GPT-5.5 · cache read0.50—
GPT-5.5 Cyber Preview12.5075.00
GPT-5.5 Cyber Preview · cache read1.25—
GPT-5.5 Pro30.00180.00
GPT-5.6 Cyber12.5075.00
GPT-5.6 Cyber · cache read1.25—
GPT-5.6 Cyber · cache write15.625—
GPT-5.6 Luna0.201.20
GPT-5.6 Luna · cache read0.02—
GPT-5.6 Luna · cache write0.25—
GPT-5.6 Sol4.0020.00
GPT-5.6 Sol · cache read0.40—
GPT-5.6 Sol · cache write5.00—
GPT-5.6 Terra2.0012.00
GPT-5.6 Terra · cache read0.20—
GPT-5.6 Terra · cache write2.50—
GPT-6 Astra10.0050.00
GPT-6 Astra · cache read1.00—
GPT-6 Astra · cache write12.50—
GPT-6 Astra Law10.0050.00
GPT-6 Astra Law · cache read1.00—
GPT-6 Cyber20.00100.00
GPT-6 Cyber · cache read2.00—
GPT-6 Cyber · cache write25.00—
GPT-6 Luna0.100.50
GPT-6 Luna · cache read0.01—
GPT-6 Luna · cache write0.125—
GPT-6 Rosalind15.0075.00
GPT-6 Rosalind · cache read1.50—
GPT-6 Rosalind · cache write18.75—
GPT-6 Sol2.0010.00
GPT-6 Sol · cache read0.20—
GPT-6 Sol · cache write2.50—
GPT-6.1 Sol2.0010.00
GPT-6.1 Sol · cache read0.10—
GPT-6.1 Sol · cache write2.50—
GPT-6.1 Terra1.005.00
GPT-6.1 Terra · cache read0.05—
GPT-6.1 Terra · cache write1.25—
GPT-rosalind Discovery5.0025.00
GPT-rosalind Discovery · cache read0.50—
GPT-rosalind Research5.0025.00
GPT-rosalind Research · cache read0.50—
o32.008.00
o3 · cache read0.50—
o3-pro20.0080.00
text-embedding-3-large0.13—
text-embedding-3-small0.02—
text-embedding-ada-0020.10—

Standard short-context text rates; cache read is the cached-input rate and cache write is the short-context 30-minute cache-write rate, on the models billed for one. Long-context prompts, processing tiers, and regional processing may have different rates. Spend uses OpenAI's billed amounts.

xAI (Grok)

ModelInputOutput
Grok 4.20 0309 Non Reasoning1.252.50
Grok 4.20 0309 Non Reasoning · cache read0.20—
Grok 4.20 0309 Reasoning1.252.50
Grok 4.20 0309 Reasoning · cache read0.20—
Grok 4.20 Multi Agent 03091.252.50
Grok 4.20 Multi Agent 0309 · cache read0.20—
Grok 4.31.252.50
Grok 4.3 · cache read0.20—
Grok 4.52.006.00
Grok 4.5 · cache read0.30—
Grok 4.62.006.00
Grok 4.6 · cache read0.50—
Grok 4.72.006.00
Grok 4.7 · cache read0.50—
Grok Build 0.11.002.00
Grok Build 0.1 · cache read0.20—

Rates shown are for prompts up to 200k tokens. Longer prompts are billed at twice these rates.

Price changes

Every price change the weekly check recorded, newest first, since .

DateProviderModelChangeBeforeAfter
OpenAIGPT-5.5 Cyber PreviewNew model—12.50 / 75.00
Anthropic (Claude)Claude Mythos 5.2New model—8.00 / 40.00
Anthropic (Claude)Claude Fable 5.2New model—8.00 / 40.00
OpenAIGPT-6 Astra LawNew model—10.00 / 50.00
OpenAIGPT-6 CyberNew model—20.00 / 100.00
OpenAIGPT-6 RosalindNew model—15.00 / 75.00
OpenAIGPT-rosalind DiscoveryNew model—5.00 / 25.00
OpenAIGPT-rosalind ResearchNew model—5.00 / 25.00
OpenAIGPT-6.1 TerraNew model—1.00 / 5.00
Anthropic (Claude)Claude Sonnet 4.5Retired3.00 / 15.00—

Reading these prices

A list price is where your bill starts, not what it ends at. Two models at the same price per token can cost very different amounts for the same text, because each one counts tokens with its own tokenizer. The token cost calculator shows that difference on your own text.

Cache rows matter more than they look. A cache read usually costs a fraction of the input price, so a long conversation that reuses its context can bill well below the headline rate. A cache write can cost more than plain input. Rows marked off peak apply only during the hours the provider names.

Discounts, batch rates, regional pricing and credits are not in these tables. Your provider's invoice is the final word on what you were charged.

See what these prices add up to

Connect your providers and CostCompass shows your month-to-date spend across all of them, with a forecast for next month.

Start free