LLM API prices per million tokens
List prices for standard text tokens, in USD per million tokens. “Same text” is the input price per million tokens of text as GPT and Gemini count it, so it compares like with like across tokenizers.
Free tool · Prices checked October 19, 2026
LLM pricing comparison
Price lists compare tokens. You pay for tasks. Pick a workload and see what it costs a month on 12 current models from Anthropic, OpenAI and Google, with prompt caching, long-context prices and each provider's tokenizer counted.
- cost spread for the same support agent
- 137×
- cost spread for the same support agent
- of an agent's bill is input, not output
- 88%
- of an agent's bill is input, not output
- the tokens on Claude for the same text
- 1.6×
- the tokens on Claude for the same text
Answers a billing question in 5 calls: account, invoice, refund status, then the reply.
6K system prompt · 5 calls · 2K per tool result · 300 out per call
GPT-6 Luna is cheapest at $71.13 a month. Claude Fable 5.1 costs $5,712, 80× more for the same app.
GPT-6 Luna$71.13 a month$0.0024 each
Claude Haiku 5.5$109 a monthinstead of$114
Gemini 3.1 Flash-Lite$154 a monthinstead of$171- Gemini 3.5 Flash-Lite$209 a monthinstead of
$237 - Gemini 3.8 Flash$349 a monthinstead of
$481 - Gemini 4 Argon$760 a monthinstead of
$1,148 - GPT-6.1 Sol$801 a monthinstead of
$1,289 - Gemini 3.1 Pro Preview$954 a monthinstead of
$1,372 - Claude Sonnet 5.5$1,258 a monthinstead of
$2,062 - Claude Opus 5.5$2,475 a monthinstead of
$4,124 - GPT-6 Astra$4,364 a monthinstead of
$7,113 - Claude Fable 5.1$5,712 a monthinstead of
$9,775
AnthropicOpenAIGoogleWith Contextwaise, grey: before. 1 model unchanged: the plan would cost more than it saves.What each call sends
Every call re-sends the whole task so far. Tokens per call, as GPT and Gemini count them.
- Read from cache
- Billed at full input price
- Output
- Removed by Contextwaise
- Long-context price starts
List prices for standard (non-batch) text tokens. Caching assumes the shared system prompt stays warm; on Gemini it is implicit caching, whose hits are likely but not guaranteed. Contextwaise applies the cuts measured on the demo (context −65%, tool output −83%) and adds the cheapest plan for the tokens it reads. Your results will differ: measure on your own traffic.
| Model | Input | Cached input | Output | Input, same text | Long context |
|---|---|---|---|---|---|
| Anthropic | |||||
| Claude Fable 5.1 | $10.00 | $0.25 | $50.00 | $16.00 | Same price |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | $6.40 | Same price |
| Claude Sonnet 5.5 | $2.00 | $0.10 | $10.00 | $3.20 | Same price |
| Claude Haiku 5.5 | $0.10 | $0.01 | $0.50 | $0.16 | $0.50 / $2.50 above 100K |
| OpenAI | |||||
| GPT-6 Astra | $10.00 | $1.00 | $50.00 | $10.00 | $20.00 / $75.00 above 272K |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $2.00 | $4.00 / $15.00 above 272K |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 | $0.10 | $0.20 / $0.75 above 272K |
| Google | |||||
| Gemini 4 Argon | $2.00 | $0.10 | $10.00 | $2.00 | Same price |
| Gemini 3.1 Pro Preview | $2.00 | $0.20 | $12.00 | $2.00 | $4.00 / $18.00 above 200K |
| Gemini 3.8 Flash | $0.75 | $0.075 | $3.75 | $0.75 | Same price |
| Gemini 3.5 Flash-Lite | $0.30 | $0.03 | $2.50 | $0.30 | Same price |
| Gemini 3.1 Flash-Lite | $0.25 | $0.025 | $1.50 | $0.25 | Same price |
Sources: Anthropic pricing (checked October 14, 2026) · OpenAI pricing (checked October 19, 2026) · Google pricing (checked October 14, 2026). Batch APIs halve input and output prices on all three. Provider pages: Claude API pricing, Gemini API pricing.
Announced price changes
- Gemini 4 Argon: $2.00 → $4.00 input and $10.00 → $20.00 output from December 1, 2026.
- Gemini 3.8 Flash: $0.75 → $1.50 input and $3.75 → $7.50 output from January 1, 2027.
Why an agent's bill grows with every call
A price list tells you what one token costs. The bill depends on how many tokens your app sends, and an agent sends far more than it seems. Here is the support agent from the tool, one conversation, call by call:
Call 1
6.2K
Call 2
8.5K
Call 3
10.8K
Call 4
13.1K
Call 5
15.4K
- System prompt and tools
- Question
- Earlier calls, re-sent
- Latest answer and tool result
- Output
On GPT-6.1 Sol without caching, call 1 costs $0.0154 and call 5 costs $0.0338: the same question gets 2.2× more expensive by its last step. Right column: input tokens per call.
- 1The system prompt is billed on every call. Instructions and tool definitions, 6K tokens, travel with each of the 5 calls: 30K tokens per conversation for the same text.
- 2Tool results pile up. Each tool returns 2K tokens of JSON. They stay in the history, so the result from call 1 is billed again in calls 2 to 5.
- 3The answer is the small part. The model writes 300 tokens per call. Output costs 5× to 8.3× more per token, yet input is 88% of this bill.
- 4Caching turns repeats into discounts. Everything a call shares with the one before is a repeated prefix. With caching on, it is billed at 2.5% to 10% of the input price, which is why the tool's bars drop by more than half.
Count your own prompt with the token counter, and see the LLM cost optimization guide for the levers that shrink each part.
Which LLM is cheapest? It depends on the app
The cheapest model per token is not always the cheapest per task. Rankings move with the workload, because three rules sit on top of the price list:
- Cache read prices differ more than list prices. Claude Fable 5.1 and GPT-6 Astra both list $10.00 input and $50.00 output, but Fable reads its cache at $0.25 and Astra at $1.00. On the 60-call coding agent, where almost everything is a cache read, Fable costs $9,226 a month against $10,811.
- Long-context prices kick in mid-task. Claude Haiku 5.5 bills 5× its price past 100K tokens. The coding agent crosses that at call 16, so Haiku drops from #2 on the support agent to #4 here.
- Short prompts don't cache everywhere. The classifier's 600-token prompt is under the GPT and Gemini caching minimums, but over Claude's 512, so Claude Haiku 5.5 comes out cheapest despite its tokenizer.
| Workload | Cheapest | Cheapest of each provider | Spread |
|---|---|---|---|
Support agent 30,000 conversations a month | GPT-6 Luna $71.13 a month | Claude Haiku 5.5 $114GPT-6 Luna $71.13Gemini 3.1 Flash-Lite $171 | 137× |
RAG chatbot 100,000 questions a month | GPT-6 Luna $103 a month | Claude Haiku 5.5 $164GPT-6 Luna $103Gemini 3.1 Flash-Lite $300 | 158× |
Coding agent 1,000 tasks a month | GPT-6 Luna $108 a month | Claude Haiku 5.5 $769GPT-6 Luna $108Gemini 3.1 Flash-Lite $265 | 100× |
Classifier 1,000,000 tickets a month | Claude Haiku 5.5 $49.60 a month | Claude Haiku 5.5 $49.60GPT-6 Luna $85.00Gemini 3.1 Flash-Lite $215 | 171× |
Prompt caching on, Claude billed with its own tokenizer, today's prices. Change any of it in the tool above.
Price is half the decision. Route each request type to the cheapest model that passes your own evaluation set, and keep the expensive model for the requests that need it.
How this comparison works
- Workloads in tokens of text. Sizes are counted the way GPT and Gemini count (about 4 characters a token). With “Claude's tokenizer” on, Claude is billed for 1.6× the tokens, from Anthropic's published ratio.
- Stateless calls. Each call re-sends the system prompt, the question and every earlier answer and tool result.
- Caching. A cache marker ends every call of a multi-call task, so each call reads the previous one from the cache and writes the rest. One-call tasks cache the system prompt. The shared system prompt is assumed warm. Prefixes under a provider's minimum are not cached.
- Long context. Each call is priced at the tier its size falls in.
- Contextwaise. Applies the cuts measured on the demo (retrieved context −65%, tool output −83%) and adds the cheapest plan for the tokens it reads. On cheap models it can cost more than it saves, and the tool shows that.
- Not included. Batch discounts, Gemini's explicit cache storage, web search and other tool fees, images and audio.
Frequently asked questions
Which LLM API is the cheapest?
On list price, Claude Haiku 5.5 ($0.10 per million input tokens, $0.50 output) and GPT-6 Luna ($0.10, $0.50) are the cheapest current models from Anthropic, OpenAI and Google. Per task it depends on the workload: Claude's tokenizer counts more tokens for the same text, while its shorter caching minimum lets small prompts cache. In the comparison above, GPT-6 Luna is cheapest for the support agent and Claude Haiku 5.5 for the classifier.
How much does an LLM API cost per month?
It depends on the model more than anything else. The same support agent at 30,000 conversations a month costs $71.13 on GPT-6 Luna and $9,775 on Claude Fable 5.1, with prompt caching on. Multiply your tokens per task by your tasks per month, then by the model's prices, or pick the closest workload in the tool above.
Why does the same prompt cost a different number of tokens on Claude, GPT and Gemini?
Each provider trains its own tokenizer. Anthropic says 1M tokens is about 2.5M characters on current Claude models; OpenAI and Google say a token is about 4 characters. The same English text is about 1.6× the tokens on Claude, so compare prices per task, not per token.
Does prompt caching work the same on every provider?
No. Anthropic and OpenAI's GPT-6 models bill cache writes at 1.3× the input price and cache reads at 2.5% to 10% of it. Gemini's implicit caching has no write fee, but hits are not guaranteed. The shortest prefix that caches also differs: 512 tokens on Anthropic, 1,024 tokens on OpenAI, 4,096 tokens on Google.
What is long-context pricing?
Some models charge more per token once a prompt passes a size: Claude Haiku 5.5 above 100K input tokens, Gemini 3.1 Pro Preview above 200K input tokens, OpenAI models above 272K input tokens. The higher price applies to the whole request, input and output. Agents that keep every tool result in the context cross these lines after enough steps.
Are output tokens more expensive than input tokens?
Yes, 5× to 8.3× more per token on current models. But input is usually most of the bill: in a 5-call agent, 88% of the cost is input, because every call re-sends the conversation so far.
How often are these prices updated?
Prices come from each provider's pricing page: Anthropic checked October 14, 2026, OpenAI checked October 19, 2026, Google checked October 14, 2026. Announced changes are in the table, and the "Announced prices" switch applies them.
Cut your LLM costs
without cutting quality.
One layer between retrieval and inference removes the context your model never needed, in milliseconds. Fewer input tokens, lower cost, a prompt cache that keeps hitting, and every removal logged with its reason.
Keep your model, retriever and prompts. Remove one call to roll back.