LIST TARIFF List prices and the arithmetic behind the bill
Rates verified 18 November 2026 22 models 3 providers USD per 1,000,000 tokens Source: vendor price pages only

Calculator 01 — token spend

What one month of API calls actually costs

Enter the shape of your workload. The calculator applies each provider's published input, output, cache and batch rates — including the thresholds where the price per token changes — and prints the arithmetic it used.

Prices: standard tier Taxes and negotiated discounts: excluded Next scheduled change: 1 Dec 2026
01Workload
One API call, one request.
Prompt, system text, history, retrieved context.
Visible output plus reasoning tokens where billed as output.
Share of input tokens served from a warm cache.
0 in steady state. Billed at the shortest cache-write rate.
Standard on-demand rates.
02Result
Model Provider Cost / request Cost / month Effective / 1M vs lowest

Sorted by monthly cost. Select any row to see its arithmetic below. Effective rate = monthly cost ÷ all tokens billed.

03Rate table

List prices, per 1M tokens

Standard tier, text tokens, USD. Figures as published on 18 November 2026. Footnotes mark the exceptions the calculator applies.

Model Input Output Cached read Batch in / out Billing note
OpenAI
GPT-6.2 Astra8.0040.000.404.00 / 20.00>272K input: 16.00 / 60.001
GPT-6.2 Sol2.0010.000.101.00 / 5.00>272K input: 4.00 / 15.001
GPT-6 Astra10.0050.001.005.00 / 25.00>272K input: 20.00 / 75.001
GPT-6.1 Sol2.0010.000.101.00 / 5.00>272K input: 4.00 / 15.001
GPT-6.1 Terra1.005.000.050.50 / 2.50>272K input: 2.00 / 7.501
GPT-6 Luna0.100.500.010.05 / 0.25>272K input: 0.20 / 0.751
GPT-6 Sol2.0010.000.201.00 / 5.00>272K input: 4.00 / 15.001
GPT-5.6 Sol4.0020.000.402.00 / 10.00Promotional price until 30 Nov 20265
GPT-5.6 Terra2.0012.000.201.00 / 6.00>272K input: 4.00 / 18.001
GPT-5.6 Luna0.201.200.020.10 / 0.60>272K input: 0.40 / 1.801
GPT-5.4 Mini0.754.500.0750.375 / 2.25Single tier
GPT-5 Nano0.050.400.0050.025 / 0.20Single tier
Anthropic
Claude Fable 5.28.0040.000.204.00 / 20.001M context at standard rate2
Claude Opus 5.54.0020.000.202.00 / 10.001M context at standard rate2
Claude Sonnet 5.52.0010.000.101.00 / 5.001M context at standard rate2
Claude Sonnet 52.0010.000.201.00 / 5.001M context at standard rate2
Claude Haiku 5.50.100.500.010.05 / 0.25>100K prompt: 0.50 / 2.503
Claude Haiku 4.51.005.000.100.50 / 2.50Single tier
Google
Gemini 4 Carbon2.0010.000.101.00 / 5.006Doubles on 1 Dec 20264
Gemini 4 Argon2.0010.000.101.00 / 5.006Doubles on 1 Dec 20264
Gemini 4 Hydrogen1.507.500.0750.75 / 3.756Single tier, free tier available
Gemini 3.8 Flash0.753.750.0750.375 / 1.8756Doubles on 1 Jan 20274
1 · OpenAI long context
A request above 272,000 input tokens is billed end to end at the long-context rates in the note. Cache reads follow the long-context cached rate.
2 · Anthropic context
Claude 4.6 and later, except Haiku 5.5, include the full 1M-token window at standard rates. Prompt caching and batch discounts apply across it.
3 · Haiku 5.5 tier
A prompt above 100,000 tokens — counting cache reads and writes — is billed at 0.50 input / 2.50 output for that request.
4 · Google increase
Gemini 4 Carbon and Argon list at 2.00 / 10.00 through 30 November 2026 and 4.00 / 20.00 from 1 December 2026. Gemini 3.8 Flash moves to 1.50 / 7.50 on 1 January 2027.
5 · OpenAI promotion
GPT-5.6 Sol promotional pricing is available until 30 November 2026 per the price page.
6 · Google batch
Google states a 50% Batch API reduction for paid tier; per-model batch prices are not listed, so these cells are computed at that factor.
04Billing rules

Eight rules that move the total

  1. Batch is half price on both sides. OpenAI and Anthropic publish batch columns at 50% of standard input and output; Google states a 50% Batch API reduction. Batch is asynchronous, so anything user-facing cannot use it.
  2. Cache reads are billed, at a different rate. Anthropic charges 0.1× base input per read, 0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.2. OpenAI lists a per-model cached-input rate — 5% of input on the 6.x line, 10% on GPT-6 Sol.
  3. Filling a cache costs more than plain input. Anthropic writes at 1.25× base for a 5-minute cache and 2× for one hour. OpenAI lists separate 30-minute and 12-hour write rates, for example 2.50 and 5.00 on GPT-6.2 Sol against 2.00 input.
  4. A cache pays off quickly, then only if it is re-read. Anthropic's own break-even: one read for a 5-minute cache, two reads for a 1-hour cache. A write that is never read is pure loss.
  5. Long context reprices the entire request, not the excess. On OpenAI a prompt over 272K tokens bills every token at the long rate (GPT-6.2 Sol: 4.00 / 15.00 instead of 2.00 / 10.00).
  6. Thinking tokens are output tokens. Google states the output price includes thinking tokens. On OpenAI and Anthropic, reasoning and extended-thinking tokens are billed as output; raise the output estimate if reasoning is on.
  7. The tokenizer changed. Claude 4.7 and later produce roughly 30% more tokens for the same text than 4.6 and earlier. Comparing generations by token count alone understates the newer model's bill.
  8. Optional routing and speed modes add a multiplier. OpenAI regional endpoints and FedRAMP carry +10% on models released from 5 March 2026; Anthropic's inference_geo:"us" multiplies all token categories by 1.1. Fast mode doubles Anthropic's list (Fable 5.2: 16.00 / 80.00) and OpenAI's Fast tier doubles the standard rate.
05Calendar

Scheduled price changes

Only changes stated on a vendor price page, with the date the page gives.

EffectiveModelChangeWhat it does to a bill
30 Nov 2026GPT-5.6 Solpromo endsReverts to list; re-price anything pinned to it.
1 Dec 2026Gemini 4 Carbon, 4 Argon2.00 → 4.00 in
10.00 → 20.00 out
A bill built on these doubles.
1 Jan 2027Gemini 3.8 Flash0.75 → 1.50 in
3.75 → 7.50 out
Doubles; batch and cache scale with it.
28 Feb 2027GPT-6.2 Astra, 6.2 Sol — Ultrafast3× → 6× standardOnly affects Ultrafast tier users.
06Sources

Where every number comes from

Three vendor price pages, read on 18 November 2026. No aggregator, blog or screenshot was used as a source.

  1. OpenAI API pricing — developers.openai.com/api/docs/pricing. Standard, Batch, Flex, Fast and Ultrafast tables; short/long context split at 272K input tokens; cached-input and cache-write columns.
    https://developers.openai.com/api/docs/pricing
  2. Anthropic Claude pricing — platform.claude.com/docs/en/about-claude/pricing. Model table, cache multipliers, batch table, long-context rules, fast-mode and residency multipliers, tokenizer note.
    https://platform.claude.com/docs/en/about-claude/pricing
  3. Google Gemini Developer API pricing — ai.google.dev/gemini-api/docs/pricing. Per-model input, output, context caching, free tier and the dated price steps.
    https://ai.google.dev/gemini-api/docs/pricing

Prices change without notice and vary by platform: Azure, Bedrock and Google Cloud resell the same models at their own rates. This page covers first-party APIs only.

07Method

What the number includes, and what it does not

Included
List price per 1M tokens for text input and output, cached reads, cache writes at the shortest published duration, batch pricing, and the long-context tier when your input crosses the threshold.
Excluded
Negotiated or committed-use discounts, provisioned throughput, taxes, and charges outside token pricing — web search calls, code-execution hours, grounding, image or audio tokens.
Token counts
You supply them. No tokenizer runs in this page; count with the provider's own tokenizer, and remember images, PDFs and tool schemas all consume input tokens.
Mixed cache and batch
No vendor publishes a batch cache-read rate. With batch on, the cache-read line is the list cache rate × 0.5 and is labelled an estimate.
08FAQ

Questions this page answers

How does the calculator work out a monthly API bill?
It takes requests per month, input and output tokens per request, the cache-read share and the cache-write volume, applies each provider's published input, output, cache and batch rates — including the thresholds where the price per token changes — and prints the arithmetic for the selected model: cost per request, cost per month, effective rate per 1M tokens and the gap to the cheapest model in the table.
Are batch rates and cache reads included?
Yes. OpenAI and Anthropic publish batch columns at 50% of standard input and output, and Google states a 50% Batch API reduction; batch is asynchronous, so anything user-facing cannot use it. Cache reads are billed at a different rate: Anthropic charges 0.1× base input per read (0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.2), and OpenAI lists a per-model cached-input rate — 5% of input on the 6.x line, 10% on GPT-6 Sol.
What happens above the long-context threshold?
On OpenAI a prompt over 272K input tokens bills every token of that request at the long-context rate, not just the excess: GPT-6.2 Sol moves from 2.00 / 10.00 to 4.00 / 15.00 per 1M tokens. Claude 4.6 and later, except Haiku 5.5, include the full 1M-token window at standard rates, while a Haiku 5.5 prompt above 100,000 tokens is billed at 0.50 input / 2.50 output for that request.
Which prices are excluded, and which are about to change?
The figures are standard-tier list prices as published on 18 November 2026. Taxes, negotiated discounts and third-party platforms such as Azure, Bedrock and Google Cloud are excluded. Scheduled changes are dated on the page: Gemini 4 Carbon and Argon move from 2.00 / 10.00 to 4.00 / 20.00 on 1 December 2026, Gemini 3.8 Flash from 0.75 / 3.75 to 1.50 / 7.50 on 1 January 2027, and GPT-5.6 Sol's promotional price ends on 30 November 2026.