Guide · 5 September 2026
LLM API pricing compared (September 2026): what a real request costs on GPT, Claude and Gemini
Thirteen current OpenAI, Anthropic and Google models priced per million tokens and, more usefully, per request and per month. Verified against official pricing pages, with the caching, batch and long-context rules that change the answer.
Every provider publishes a price per million tokens, and every comparison article copies those numbers into a table. It’s the right starting point and the wrong place to stop, because nobody buys a million tokens of anything. You buy requests — each with a prompt of a certain size, a reply of a certain size, and a tokenizer that decides how big “a certain size” is. This guide gives you the list prices (verified 5 September 2026 against each provider’s official page), then shows how to turn them into the number you actually need: dollars per month for your feature.
List prices, standard tier
Per million tokens, real-time, uncached, prompts up to 200K tokens. Promotional rates marked ◦.
| Model | Input | Output | Context | Notes |
|---|---|---|---|---|
| OpenAI | Tokenizer: o200k (public) | |||
| GPT-6 Astra | $10.00 | $50.00 | ~1M | Higher long-context tier above 200K |
| GPT-5.6 Sol ◦ | $4.00 | $20.00 | 1.05M | Promotional at least through 21 Nov 2026 |
| GPT-5.6 Terra | $2.00 | $12.00 | 1.05M | |
| GPT-5.6 Luna | $0.20 | $1.20 | 1.05M | |
| GPT-5.4 mini | $0.75 | $4.50 | 400K | |
| GPT-4o mini | $0.15 | $0.60 | 128K | Cheapest current OpenAI model |
| Anthropic | Tokenizer: proprietary | |||
| Claude Fable 5.1 | $10.00 | $50.00 | 200K | Cache reads at 2.5% of input |
| Claude Opus 5 | $5.00 | $25.00 | 200K | |
| Claude Sonnet 5 | $2.00 | $10.00 | 200K | $2/$10 made permanent Aug 2026 |
| Claude Haiku 4.5 | $1.00 | $5.00 | 200K | |
| Tokenizer: proprietary | ||||
| Gemini 3.1 Pro | $2.00 | $12.00 | 1M | $4 / $18 above 200K tokens |
| Gemini 3.8 Flash ◦ | $0.75 | $3.75 | 1M | Promotional through 31 Dec 2026 |
| Gemini 2.5 Flash | $0.30 | $2.50 | 1M |
Three patterns jump out. Output is 5–6× input everywhere. The flagship tier has converged on $10 / $50 across OpenAI and Anthropic. And the “mid” tier — Terra, Sonnet 5, Gemini 3.1 Pro — has converged on roughly $2 / $10–12, which makes tokenizer differences and quality the deciding factors there, not price.
From list price to cost per request
Take a realistic customer-support request: a 1,200-token system prompt, 300 tokens of conversation and user message, and a 350-token reply.
| Model | Input cost | Output cost | Per request | Output share |
|---|---|---|---|---|
| GPT-4o mini | $0.000225 | $0.000210 | $0.00044 | 48% |
| GPT-5.6 Luna | $0.000300 | $0.000420 | $0.00072 | 58% |
| Gemini 2.5 Flash | $0.000450 | $0.000875 | $0.00133 | 66% |
| Claude Haiku 4.5 | ~$0.001725 | $0.001750 | ~$0.00348 | 50% |
| GPT-5.6 Terra | $0.003000 | $0.004200 | $0.00720 | 58% |
| Claude Sonnet 5 | ~$0.003450 | $0.003500 | ~$0.00695 | 50% |
| Gemini 3.1 Pro | $0.003000 | $0.004200 | $0.00720 | 58% |
| Claude Opus 5 | ~$0.008625 | $0.008750 | ~$0.01738 | 50% |
| GPT-6 Astra | $0.015000 | $0.017500 | $0.03250 | 54% |
Claude rows use a ~1.15× token estimate, since Anthropic’s tokenizer isn’t public; that’s why the Sonnet and Terra rows, at identical list prices for input, come out slightly different. Notice the output share column: with a 350-token reply, half to two-thirds of every request is the answer, not the prompt. Cutting the reply from 350 to 200 tokens saves more on Terra than switching from Terra to Sonnet would.
At 10,000 requests a day, multiply by 300,000: GPT-4o mini is about $130/month, Terra or Gemini Pro about $2,160, Astra about $9,750. Same feature, 75× spread. That is the decision worth spending an afternoon on, and the LLM API Cost Calculator does the arithmetic on your real prompt rather than this illustrative one — paste it, set your reply length and requests per day, and read the per-month column.
The three modifiers that move the answer
Prompt caching. If your system prompt and tool definitions are identical from request to request — they usually are — OpenAI and Anthropic bill that repeated prefix at about a tenth of the input rate on a cache hit (Anthropic: $0.20 per million on Sonnet 5 instead of $2; Fable 5.1 goes to $0.25 instead of $10). Anthropic charges a one-time write premium (1.25× input) the first time a prefix is cached; OpenAI’s caching is automatic on prompts over a minimum length. For the 1,200-token system prompt above, caching drops the cacheable input cost by roughly 90%, which takes 25–35% off the whole request. Design your prompt so the stable part comes first — caching works on prefixes.
Batch processing. Both OpenAI and Anthropic take about 50% off input and output for jobs submitted to their batch endpoints, with results returned within 24 hours. Classification backfills, nightly summaries, embedding-adjacent enrichment — anything not waiting on a human — should run there.
Long context. Above 200K tokens in a single prompt, Gemini Pro doubles its rates ($4 / $18), and OpenAI’s newest models move to a higher tier too. A million-token window is a capability, not a price plan; a RAG pipeline that occasionally overfills its context pays the premium on the whole request. Watch the fit bars in the counter, and summarise or retrieve rather than stuff.
How to choose, in practice
- Write one representative prompt for the feature — real system prompt, real tool schemas, a realistic user turn.
- Measure your typical output length from logs, or estimate: a terse classification is 5–20 tokens, a chat answer 150–400, a page of generated text 600–800.
- Paste it into the calculator, enter output tokens and requests per day, and sort by the monthly column.
- Draw a quality line. Test the two or three cheapest models on 50 real examples. The cheapest one that passes is your default; route hard cases to a stronger model only when a cheap check says they’re hard.
- Then apply the modifiers: cache the prefix, batch the offline work, cap output length in the instruction. Re-measure after a week of traffic.
A note on this table’s shelf life
Prices in this table were read from the providers’ official pages on 5 September 2026. Two are explicitly promotional with end dates, one changed direction in August, and all three providers have moved prices more than once this year. The calculator shows its own verification date under the table; if that date is more than a couple of months old, check the official pages before signing anything.
Checklist
- Compare cost per request on your prompt, not price per million tokens.
- Expect output to be half or more of each request’s cost; shorten replies first.
- Put the stable part of the prompt first and turn on caching.
- Send anything that can wait to the batch endpoint for ~50% off.
- Stay under 200K tokens per request unless the long-context rate is in the budget.
- Re-check prices quarterly; note promo end dates (Sol: Nov 2026, Gemini 3.8 Flash: Dec 2026).
Frequently asked questions.
Which LLM API is cheapest in 2026?
On list price, GPT-4o mini ($0.15 in / $0.60 out per million tokens) and GPT-5.6 Luna ($0.20 / $1.20) are the cheapest current models from the big three, followed by Gemini 2.5 Flash ($0.30 / $2.50) and Gemini 3.8 Flash ($0.75 / $3.75, promotional). Among Claude models, Haiku 4.5 is cheapest at $1 / $5. Whether 'cheapest' holds for your workload depends on tokenizer differences and output length — run your own prompt through the calculator.
Why is output so much more expensive than input?
Generating tokens is sequential and compute-heavy; reading them is parallel and cheap. Providers price accordingly: output is 5× input on Claude and Gemini, 6× on GPT-5.6. For a typical request with a few hundred tokens of reply, output is usually the larger share of the bill.
What is prompt caching and how much does it save?
If the start of your prompt (system message, tool definitions, reference documents) is identical across requests, OpenAI and Anthropic bill that repeated prefix at about 10% of the normal input rate — 2.5% on Claude Fable 5.1. For applications with a large, stable system prompt, this cuts effective input cost by 60–90%.
What is the batch discount?
OpenAI and Anthropic offer roughly 50% off both input and output for requests submitted asynchronously and returned within 24 hours. It suits nightly classification, summarisation and enrichment jobs — anything that isn't waiting on a user.
How often do these prices change?
Several times a year. In 2026 alone, Anthropic cancelled a planned Sonnet increase, OpenAI introduced promotional pricing on GPT-5.6 Sol through November, and Google set Gemini 3.8 Flash promotional pricing through December. Treat any table as a snapshot with a date on it — including this one.
Sources
Tools for this job.
Related guides.
- How many tokens is my prompt? GPT, Claude and Gemini counted on the same text
The four-characters-per-token rule breaks budgets. How each provider actually tokenizes English, code, JSON and other languages, why the counts differ, and how to get an exact number without sending your prompt anywhere.