FUTURE PROOF MARKETER/AI-TOOLS · CALCULATOR · PRICES CHECKED 2026-09-23
AI API Cost Calculator: OpenAI, Claude and Gemini API Pricing (September 2026)
The OpenAI API costs $2 per million input tokens and $10 per million output tokens on GPT-6 Sol, the middle of its three current models; GPT-6 Astra is $10 / $50 and GPT-6 Luna $0.10 / $0.50. You pay for every token you send (input) and every token the model writes back (output), per million, with no fee per request. A typical chat reply of 1,500 tokens in and 400 out costs about $0.007. Claude and Gemini bill the same way; the calculator below does the arithmetic for 10 models.
Calculate your cost →See the price table
STANDARD API PRICES IN USD PER 1M TOKENS · FROM EACH PROVIDER'S OWN PRICING PAGE · CHECKED 2026-09-23
Cost per request, per day, per month
Pick a model, say how big a request is and how many you send. Not sure of your token counts? Start from one of the typical requests and adjust once you have real numbers from your provider's usage dashboard.
Cost per request
$0.007
Cost per day
$7.00
Cost per month
$210.00
With 20% fewer input tokens: $192.00 a month, $18.00 less (9% of the bill, because input is 43% of what this workload costs on GPT-6 Sol). Input is where agents spend: every turn resends the whole conversation, so trimming what goes in moves the bill more than trimming what comes out.
Cut the input without touching your code
Gate AI sits between your tools and the provider and removes tokens the model has already seen: repeated files become pointers, tool schemas are minified, duplicate log lines go. Gate's published typical figure on agent requests is the 20% above. Free records, Pro blocks: the free plan compresses and logs 20,000 requests a month; Pro at $20 per user per month adds prompt-injection blocking, redaction and spend caps per key (checked September 2026).
Some links below are affiliate links: if you buy through them we may earn a commission at no extra cost to you. It funds the testing budget and never changes a verdict — affiliate policy.
The same workload on every model
1,500 in / 400 out × 1,000 requests a day, per month. Cheapest first.
API prices per million tokens
Standard (not batch) prices in USD, checked 23 September 2026 on each provider's pricing page. Cached input is the price of a cache hit. Where the price depends on prompt length, the table shows the base tier and the note gives the rest.
| Model | Input | Cached input | Output | Note |
|---|---|---|---|---|
| OpenAI · developers.openai.com/api/docs/pricing | ||||
| GPT-6 Astragpt-6-astra | $10 | $1 | $50 | Short-context rate; long-context prompts cost $20 in / $75 out. |
| GPT-6 Solgpt-6-sol | $2 | $0.20 | $10 | Short-context rate; long-context prompts cost $4 in / $15 out. |
| GPT-6 Lunagpt-6-luna | $0.10 | $0.01 | $0.50 | Short-context rate; long-context prompts cost $0.20 in / $0.75 out. |
| Anthropic · platform.claude.com/docs/en/about-claude/pricing | ||||
| Claude Fable 5.1 | $10 | $0.25 | $50 | Cache writes cost $12.50 (5-minute) or $20 (1-hour). |
| Claude Opus 5.5 | $4 | $0.20 | $20 | Cache writes cost $5 (5-minute) or $8 (1-hour). |
| Claude Sonnet 5 | $2 | $0.20 | $10 | The $2 / $10 launch price is now the standard price. |
| Claude Haiku 4.5 | $1 | $0.10 | $5 | |
| Google · ai.google.dev/gemini-api/docs/pricing | ||||
| Gemini 3.1 Pro (preview)gemini-3.1-pro-preview | $2 | $0.20 | $12 | Prompts up to 200k tokens; above that $4 in / $18 out. Caching also bills storage per hour. |
| Gemini 3.8 Flashgemini-3.8-flash | $0.75 | $0.075 | $3.75 | Price through 31 Dec 2026; from 1 Jan 2027 $1.50 in / $7.50 out. |
| Gemini 3.5 Flash-Litegemini-3.5-flash-lite | $0.30 | $0.03 | $2.50 | |
Flagship, mid and small model of each family; Anthropic's two top tiers are both listed because Opus 5.5 costs less than half of Fable 5.1. Not included: data-residency surcharges (10% on OpenAI and on Claude's US-only inference), fast or priority modes, and tool fees such as web search. Prices change without notice; we re-check monthly and date every change.
How API tokens are billed
Every request has two meters. Input tokens are everything you send: the system prompt, the conversation so far, any files or search results, the tool definitions. Output tokens are everything the model writes back, including the hidden reasoning tokens thinking models produce. Output is priced five to eight times higher than input on every model in the table.
Prices are quoted per million tokens (MTok). A token is roughly four characters of English. There is no charge per request as such, which is why the request count on its own tells you little: a thousand short questions can cost less than ten long agent turns.
Why agents cost more than chat
A model has no memory between requests. To continue a conversation, your tool resends the whole thing, every turn: the instructions, every earlier message, every file it read, every tool result. Turn one might be 3,000 input tokens; turn forty of a coding session is often 60,000 or more, for a reply of a few hundred.
That is why coding agents like Claude Code, Codex and Cursor burn through budgets and plan limits: the bill is dominated by input that the model has already seen. It is also where cutting input tokens pays most, which is what the 20% line in the calculator shows.
Caching: the cheapest input is the one you already sent
All three providers discount input they have seen recently. A cache hit costs a tenth of the normal input price on most models (less on Claude Opus 5.5 and Fable 5.1). Writing to the cache can cost extra: Anthropic charges 1.25× the input price for a five-minute cache and 2× for an hour, OpenAI lists a cache-write price at 1.25×, and Google bills cache storage per hour.
Caching works when the start of every request is identical: a fixed system prompt, a fixed document, the earlier turns of a conversation. Put what changes at the end.
Batch discounts
If an answer can wait, ask for it later. OpenAI, Anthropic and Google all publish a batch rate at half the standard price for the models in this table: you submit a file of requests and get results back within hours. Good for classification, summaries, evaluations and anything nobody is watching live. The calculator uses standard prices; halve the result for batch.
How to cut your API bill, in order of effort
- 01Use the smallest model that passesRoute routine work (classification, extraction, short replies) to the small model of each family and keep the flagship for the hard part. The gap between tiers in the table is often 10× or more.
- 02Cache the stable prefixKeep the system prompt and reference material identical and at the front, so every request after the first reads it at a tenth of the price.
- 03Send fewer tokens per turnTrim history, stop re-sending files the model already has, minify tool schemas, drop duplicate log lines. A gateway that compresses requests does this without code changes.
- 04Cap the outputSet a max-tokens limit and ask for the format you need. Output is the expensive meter.
- 05Batch what can waitHalf price for anything that does not need an answer in seconds.
- 06Put a spend cap on every keyA loop that retries forever is the most expensive bug there is. Cap it per key before it finds you.
On a flat Claude or ChatGPT subscription the same levers buy you room under the usage limit instead of money back; see Claude Code rate limits. For routing between providers, compare the options in OpenRouter alternatives.
Related guides
Questions we actually get
What is the OpenAI API price per million tokens?→
On 23 September 2026, OpenAI's current models cost: GPT-6 Astra $10 input / $50 output, GPT-6 Sol $2 / $10, and GPT-6 Luna $0.10 / $0.50, per million tokens at standard processing. Cached input is a tenth of the input price, the Batch tier halves both, and very long prompts are billed at a higher long-context rate.
Is open ai api pricing per request or per token?→
Per token. Every request is billed as input tokens times the input price plus output tokens times the output price; there is no fee per call. That is why a long system prompt or a long conversation costs money on every single request, even when the reply is one line.
How much does the ChatGPT API cost?→
There is no separate "ChatGPT API": ChatGPT is the app, and developers use the same models through the OpenAI API, billed per token and separate from any ChatGPT Plus or Team subscription. A typical chat reply of 1,500 tokens in and 400 out costs about $0.007 on GPT-6 Sol, so 1,000 of them a day is roughly $210.00 a month.
Is there a token cost calculator for Claude?→
Yes, the calculator on this page covers Claude: pick Claude Sonnet 5, Opus or Claude Haiku 4.5, enter tokens in and out per request and requests per day. One thing to know: Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, so count tokens with the model you will actually run.
How do I use a token cost calculator for GPT when I don't know my token count?→
Estimate, then measure. A token is roughly four characters or three quarters of an English word, so 1,000 words is about 1,300 tokens. Once you have real traffic, the usage field in every API response and your provider's dashboard give exact counts; put those numbers back into the calculator.
What is the cost per request of an LLM call?→
Cost per request = (input tokens × input price + output tokens × output price) ÷ 1,000,000. For agents, input tokens include the whole conversation so far, so the cost per request climbs every turn. Multiply by requests per day and by 30 for a monthly figure; that is exactly what the calculator does.
The numbers, as they land
Our own token counts behind Gate every week from 30 September, the fifteen-minute setup checklist now, and one tested tool a week. In the letter.
One email with the goods, then the weekly letter. Unsubscribe anytime.