Sonnet 5's five lines, and the agent loop against Opus.
The rate card is read in this tab. Nothing is uploaded and no model API is called.
Last updated 2026-09-22 Source
Versus Claude Opus 5: $4.88 less than Claude Opus 5. Claude Opus 5
Sonnet 5 is the workhorse of the three Anthropic pages here. The loop is where it earns its own page. At a low hit rate you pay to write the prefix on every turn, and Sonnet's write lines are the ones a long agent session actually hits. At a high hit rate the read line dominates, and the gap versus Opus shrinks to whatever the read and output lines still differ by. That shrink is the comparison this page computes. People who only remember the input rate will overstate the savings of switching, and people who only remember the output rate will understate the cost of a cold prefix. The five lines are all on the card. None of them is a long-context tier.
The numbers on this page are not typed in. A checked rate card lives in one table, and every figure you see is read from that table or returned by the cost function the rest of the site calls. The scenario is agent-loop-50: fifty turns, one cacheable context that comes back every turn, a little new input that is never cached, and output on every turn. The three totals further down set the cache hit rate at zero, one half, and nine tenths. Hit rate is the share of that repeated context that is read from cache instead of written again. It is not a discount applied after the fact, and it is not the batch factor. Batch, when the card has one, is listed on the card and is left out of these totals on purpose, because a batch job is a different purchase from the interactive loop. When a card publishes no price for a cache line the scenario still has to account for, those tokens are billed as fresh input and the result is marked so. That is a property of the card, not a missing feature of the page. A long-context tier, when the card has one, applies to a single request whose prompt is over the threshold. The fifty short turns of this loop stay on the base tier, which is why a model with a long-context tier still shows one set of loop totals. The worked example below names the day the card was confirmed against the vendor page linked next to it. If you need a different loop, the parent page has the other three scenarios and a slider.
Rates on this page were checked against the vendor pricing page on 2026-09-22.
0% cache hit: $5.55 50% cache hit: $3.25 90% cache hit: $1.41
Versus Claude Opus 5: $4.88 less than Claude Opus 5.
It is the day this card was checked against the vendor pricing page. It is not the model's release day, and it is not a promise that the vendor has not changed a number since. When the card changes, the date changes with it.
This card publishes cache read, a five-minute cache write and a one-hour cache write, plus a batch factor. It does not publish a long-context tier, so a full window is the same five lines as a short prompt. The one-hour write is a different product from the five-minute write, not a longer spell of the same line.
Opus is the card you would move a hard loop onto. Same five lines. Read the delta at a low hit rate and at a high one before you decide the input line told you the whole story.