Claude API pricing, at half list
Every model on this page costs exactly 50% of the vendor's published list price. No tiers, no volume gates, no negotiation, no subscription. You top up prepaid credit and it is metered per token as you spend it.
The short version: Claude Opus 5 at $2.50 per million input tokens against a $5.00 list price, and Claude Sonnet 5 at $1.00 against $2.00. Same API, same model ids, same weights.
Prices in USD, billed per token from your prepaid credit. Cached input reads at 0.10× the input rate; cache writes at 1.25×. All models draw on one credit pool — there are no per-model balances. Struck-through column is the vendor's own list price for input, so you can check the halving yourself.
Previous generations still callable on the same key (6)
These are not what we would put you on for new work, but they still run and the price has not changed. Model ids in these generations are pinned snapshots, so if you validated output against one of them it keeps behaving the same way.
Per-token price is not what you pay
A rate card compares price per million tokens. Your invoice is the sum of tasks, and a more expensive model usually writes longer answers — so the real gap is wider than the rate card implies. This is the most common way a pricing comparison misleads.
We measured it rather than assert it. One identical code review — find a concurrency bug in a short Python billing function, name two lesser issues, rewrite it — sent to both models at temperature 0 with the same cap:
The gap is 3.27×, not 2.5×. Opus 5 costs 2.5 times more per token but cost 3.27 times more for the task, because it chose to emit 875 output tokens where Sonnet 5 used 592. Both found the actual bug and both flagged the same secondary issues.
Two practical consequences. Budget from cost per task, not cost per token, or you will under-forecast the expensive model by roughly a third. And on output-heavy work — code generation, long summaries, agent loops — the cheaper model saves more than its rate suggests.
Measured 31 July 2026 on our own keys; the dollar figures are what our billing system actually charged, read from our request log, not arithmetic on the rate card. One run per model, so it shows cost divergence, not model quality. Full detail and both answers are on the Opus vs Sonnet comparison.
How billing works
Pay $10, $50, $200 or any custom amount by card. Credit is added instantly. There is no monthly bill and no auto-charge, so you cannot be surprised by an invoice.
Every call is metered per token at the rate above. Remaining balance is in the dashboard
and at GET /v1/balance, so you can alert on it
from your own monitoring.
Top up again when credit runs low. Unused credit does not expire. Refunds available within 14 days per our refund policy.
Questions we get about the price
Why is it half price?
We buy upstream capacity in bulk and resell it per token with a margin, the same shape as any reseller. There is no trick in the billing: the model ids, weights and API surface are the vendor's. What you give up is the vendor's own support relationship and their enterprise contracting — if you need a DPA signed with Anthropic directly, buy from Anthropic directly.
Does cheaper mean rate-limited or queued?
No queue and no deliberate throttle. What does happen, and we would rather say it plainly, is that upstream capacity for a specific model occasionally runs short. When a model is not callable we mark it unavailable on /models/ and remove its price rather than take signups for something that will fail.
What is the cheapest way to run a lot of volume?
Prompt caching, by a wide margin. Cached reads bill at 0.10× input rate, so a repeated system prompt or a fixed codebase context costs a tenth. After that, default to the cheaper model and escalate per task rather than picking the flagship for everything — the measurement above is the argument for it.
Can I use it with my existing tooling?
It is an OpenAI-compatible endpoint plus Anthropic's own message format, so in most tools it is a base-URL and key change. There are setup notes for specific tools if you would rather copy a config than work it out.
$10 to start
Prepaid credit, no subscription, $10 minimum. Enough to run a few hundred real code reviews at the numbers above and decide with your own bill.