Claude API Pricing in 2026 (Every Model) — and How to Pay 49–60% Less

Anthropic charges per million tokens, split into input and output, and output is 5× input on every current model. Call Claude through this API instead of going direct and you pay 49–60% less per million tokens for the identical model at standard rates, from a single prepaid balance — a little more with the largest top-up pack.

Here's the full table, then how to work out what you'll actually spend.

Claude API pricing: direct vs you.bot

Prices in USD per million tokens, checked 15 August 2026.

Claude Sonnet 4.6 is priced identically to Sonnet 5. Buying credits in the $1,250 top-up pack (+10% bonus) adds a few more points on top of every row.

Model pages: Claude Sonnet 5 · Claude Opus 5 · Claude Opus 4.8 · Claude Fable 5 · Claude Haiku 4.5. Live prices for every Claude model are on the pricing page.

The newest Opus is now the cheapest Opus

This flipped recently and it's worth checking your config. Opus 5 and Opus 4.8 are $2 per million input tokens. Opus 4.7, 4.6 and 4.5 are $2.5673 — 28% more expensive for an older revision. If you pinned an older Opus for cost reasons, that reason no longer exists; move up.

Same story at the top of the range: Fable 5 dropped to $4 / $20, a 60% discount, which is the deepest cut in the Claude line-up.

What that means in real money

A support assistant. 2,000 input tokens and 400 output tokens per conversation, 50,000 conversations a month — 100M input and 20M output tokens.

A code-review bot on Opus 5. 20M input, 5M output per month:

Nothing about the model changes. Same weights, same responses, same context window — a different billing relationship.

The line item most people forget: output tokens

Output costs 5× input on every Claude tier. Which means the cheapest optimisation isn't switching models, it's making the model say less:

• Ask for JSON or a fixed schema instead of prose.
• Cap max_tokens at what you'll actually render.
• Don't ask for a summary and an explanation of the summary.

Trimming a 600-token answer to 200 on Sonnet 5 saves you more than moving from Opus to Sonnet on the input side.

Pick the right tier before you tune anything

How to call Claude through one endpoint

POST https://you.bot/api/v1/generate
Authorization: Bearer YOUR_API_KEY

{
  "modelId": "claude-sonnet-5",
  "input": { "prompt": "Summarise this changelog for a release note." }
}

Text models return their result inline in the create response, so there's nothing to poll. Swapping to Opus is a one-word change:

generate("claude-sonnet-5", {...})   # everyday work
generate("claude-opus-5", {...})     # the hard 5%

The same key also reaches GPT, Gemini, image, video and music models — see the text models or the full catalog. More on why we collapsed everything to one key in One key, every model.

Two billing details that quietly matter

You only pay for successful calls. Failed, errored or empty generations are refunded automatically, so a bad upstream day doesn't show up on your invoice. (Why we did that.)

Credits don't expire. 1 credit = $0.01, no monthly seat, no minimum, no clock on unused balance. $5 buys 500 credits to test with; larger packs add bonus credits (+5% at $500, +10% at $1,250).

Keep your existing key

The honest objection to any gateway is that it becomes a single point of failure between you and the model. So keep your Anthropic key and use it as a fallback route: send primary traffic to the cheaper endpoint, fall back to the direct API on error. You capture the savings on every successful call and give up nothing on reliability.

If you're comparing more broadly, the cheapest AI APIs in 2026 covers image, video and music too — where discounts run as high as 89%.

FAQ

How much does the Claude API cost? Direct, Claude Sonnet 5 is $3 per million input tokens and $15 output; Opus 5 and 4.8 are $5/$25; Fable 5 is $10/$50. Here the same models are $1.5367/$7.702, $2/$10 and $4/$20 respectively — 49–60% less at standard rates. Checked 15 August 2026.

Why is output more expensive than input? Generating tokens costs more compute than reading them. On every current Claude model output is priced 5× input, which is why limiting response length is the highest-leverage cost control you have.

Is Claude Sonnet 5 or Opus 5 better value? Sonnet 5 for the vast majority of production work — it's under half the price of Opus per token. Use Opus 5 for genuinely hard reasoning. Note that Opus 5 is now cheaper than Opus 4.7, so there's no cost reason to stay on the older revision.

Which Claude model is cheapest? Claude Haiku 4.5 at $0.506 per million input tokens and $2.5673 output.

Do I need a different SDK to switch models? No. It's one endpoint and one auth header; you change the modelId string.

Is there a free tier? You get 50 free credits on signup, no credit card required, and you can test any model in its in-browser playground before writing code.