LLM API pricing compared: OpenAI vs Anthropic vs Google

Three vendors, three pricing pages, three incompatible layouts. OpenAI splits by context length. Anthropic splits by cache duration. Google splits by prompt size and modality. Comparing them means normalising them first.

So here they are on one axis: US dollars per million tokens, standard tier, text only, as of 17 July 2026. Prices move; the date is the point.

Frontier models

Mid-tier

Cheap tier

What the tables do not tell you

Output costs five to six times input — and the ratio is astonishingly rigid. Look at the right-hand column. Every Anthropic model on this page prices output at exactly 5× input. Every OpenAI and Google model prices it at exactly 6×, with one exception: gpt-5.4-nano, the cheapest model here, is 6.25×.

Across three companies, four price tiers and a 50× spread in absolute cost, that ratio moves by a quarter of a multiple, once. Nobody publishes it as a policy; everyone converges on it.

This is the single most useful fact for budgeting, and no pricing page states it. It means the length of your answers, not the length of your prompts, is what you are really buying. A verbose model at $2 input can cost more in practice than a terse one at $5. If you want a smaller bill, cap max_tokens and ask for brevity before you switch vendors.

Long context is where they genuinely differ

This is the real divergence, and it is buried in footnotes.

Look closely at those multipliers. OpenAI and Google land on exactly the same ones — double the input, half again the output — from different pricing pages, neither stating it as a policy. Anthropic's docs are unusually blunt in the other direction: a 900k-token request is billed at the same per-token rate as a 9k-token request.

What that actually changes — and what it doesn't

Less than you might expect, and it is worth being precise rather than dramatic.

Gemini 3.1 Pro is less than half the price of gpt-5.6-sol on short prompts, and it stays cheaper past 200k. Against Claude Opus 4.8, which never surcharges:

So the surcharge narrows the gap without closing it. On these numbers the ranking does not flip: Gemini remains the cheapest frontier model at any length.

What changes is how much that lead is worth. Feed it long prompts and most of Gemini's price advantage evaporates — at which point a flat-rate model becomes defensible on grounds other than price: predictability, and no cliff at 200k to design around.

Whether you should send a 900k-token prompt at all is a different question. See context windows.

The lever that beats all of this: caching

Both OpenAI and Anthropic discount repeated context by roughly 90%:

If your application sends the same system prompt or the same documents on every request — which most applications do — caching is a bigger win than any vendor switch on this page. A 90% discount on the repeated portion dwarfs the 2× spread between frontier models. It is not free, though: writing to the cache costs a premium, so it only pays once you read it back. See how to cut your LLM API bill.

What an identical workload actually costs

One million input tokens and 100,000 output tokens — roughly a busy support bot's day. No caching, standard tier:

The frontier-to-floor spread is about 45×. Most production workloads do not need the top of this table, and the honest advice is unglamorous: start at the bottom, measure quality on your actual task, and move up only when it fails.

One dated change worth diarising

Claude Sonnet 5 gets more expensive on 1 September 2026 — from $2/$10 to $3/$15 per million tokens, a 50% rise. Anthropic publishes both prices on the same page today. If Sonnet 5 is in your stack, that is a budget line you can plan for now rather than discover on an invoice.

What we would actually do

• Don't shop on the headline input price. Output is 5–6× more expensive and is what varies with your prompt design.
• Cache before you switch. ~90% off repeated context beats every difference on this page.
• Match the model to the tier, not the brand. All three sell a frontier, a mid and a cheap model at broadly comparable prices. Loyalty is not a pricing strategy.
• If you send long prompts, check the surcharge before the rate — not because it flips the ranking, but because it quietly halves the gap you thought you were buying.