Claude Opus 5.5 pricing, benchmarks, API changes

Claude Opus 5.5 API pricing and token cost

Claude API pricing per million tokens (USD)

Input

Lower is cheaper. Bar height is relative to Fable 5.1, set to 100.

$02.557.510

$4Opus 5.5

$5Opus 5

$10Fable 5.1

Output

Same ratio as input: Opus 5.5 costs 40% of the Fable 5.1 rate.

$02040

$20Opus 5.5

$25Opus 5

$50Fable 5.1

Cache reads (prompt caching)

The one line where Fable 5.1 is close, $0.05 above Opus 5.5. Bar height is relative to Opus 5, set to 100.

$00.20.4

$0.20Opus 5.5

$0.50Opus 5

$0.25Fable 5.1

Opus 5.5 is cheaper than Opus 5 on every line of the price sheet, and its input and output rates are less than half of Fable 5.1's. Work that can wait gets the usual Batch API discount of half price, which brings Opus 5.5 down to $2 input and $10 output.

Per-token prices are 20% lower than on Claude Opus 5

Input goes from $5 to $4 per million tokens and output from $25 to $20, the same 20% cut on both sides. For a short request with no cache and a long answer, where output tokens make up most of the bill, that 20% is about all you save by switching.

Cache reads with Claude prompt caching cost 60% less

Cache reads drop from $0.50 to $0.20 per million tokens. Anthropic also changed the ratio: a cache read used to cost a tenth of the input price and now costs a twentieth, so cached context gets a bigger cut than the headline rate.

Writes follow the new input price. A five-minute entry costs 1.25 times the input rate and a one-hour entry twice it, which works out to $5 and $8 per million tokens, down from $6.25 and $10. The shortest prompt you can cache is 512 tokens.

This is where agents and coding tools spend most of their money, because they reread the same context on every turn. By the Claude blog's estimate, a session made up mostly of cache reads saves up to 60% on input, and most Claude Code tasks fall somewhere between that and the 20% floor.

The 40% figure assumes Claude effort drops to medium

Anthropic's 40% is the saving on typical work "at default settings", and it comes from two sources: the 20% lower price, and fewer tokens used per task. The defaults themselves moved. A request that doesn't set effort now runs at medium, while the same request on Opus 5 ran at high.

So the 40% doesn't tell you what you'll save if you pin effort to high or max. At the same effort level Opus 5.5 also tends to think more per turn than Opus 5, especially at xhigh and max. Measure cost per task on your own workload before you plan a budget around 40%.

Opus 5.5 lasts about 25% longer within Claude usage limits

Pro, Max, Team and seat-based Enterprise plans also get higher five-hour usage limits, though neither the announcement nor the docs say by how much. Subscribers can also bank a rate limit reset and use it when they choose.

The Claude blog does give a number for plans. The lower Opus 5.5 price carries over to your limits, cached context included, so the same plan goes about 25% further than it did on Opus 5. The extra cut on cache reads is an API price change only.

How Claude Opus 5.5 benchmarks against Fable 5.1

Anthropic-reported benchmark scores

GDPval-AA v2.1 (Elo)

All three models at max effort. Bar height is relative to Opus 5.5, set to 100.

1846Opus 5.5

1735Fable 5.1

1708Opus 5

05001,0001,500

CursorBench (%)

Opus 5.5 runs at its default medium effort; the other two run at max.

52.5Opus 5.5 (medium)

51.8Fable 5.1 (max)

46.6Opus 5 (max)

02040

FrontierCode (%)

Opus 5.5 at medium effort against the best GPT-6 Astra score.

54.6Opus 5.5 (medium)

53.3GPT-6 Astra

02040

Anthropic says Opus 5.5 leads its own benchmarks in agentic coding, computer use and knowledge work. Unless a result says otherwise, Anthropic ran each test at max effort with adaptive thinking.

Claude Opus 5.5 beats Fable 5.1 on GDPval-AA and CursorBench

GDPval-AA v2.1 grades real work across 44 occupations. With all three models at max effort, Opus 5.5 comes out on top.

The coding results are more striking because the settings favor the older models. On CursorBench, Opus 5.5 at its default medium effort beats Fable 5.1 and Opus 5 at max. On FrontierCode, medium-effort Opus 5.5 edges past GPT-6 Astra's best score for about a fifth of the cost per task.

In one early-access test, Opus 5.5 audited and repaired a 200,000-line codebase in less than 3 hours, a job that ran over 20 hours on Opus 5 and burned 2.5 times as many tokens. Anthropic also reports output speeds more than 30% above Opus 5.

Anthropic says the real gap with Claude Fable 5.1 is smaller

Anthropic qualifies its own numbers. At this level of capability, it says, benchmark margins say less about real-world differences, and inside Anthropic the two models feel closer in day-to-day use than the scores imply.

Read the chart as "Opus 5.5 is in Fable's range", not as proof that Opus 5.5 is the stronger model. The difference you will notice is the bill: Fable 5.1 costs $10 and $50 per million tokens.

Four breaking changes when migrating from Claude Opus 5

Four requests that work on Opus 5 but are rejected on Opus 5.5

Rejected with a 400 error	Use instead
thinking: disabled, or budget_tokens	Adaptive thinking with effort
tool_choice any or tool	auto or none
Old thinking blocks resent after editing system, tools or earlier messages	Append to the conversation instead of editing it
computer_20251124 on the Claude API and Google Cloud	computer_toolset_20260801

If your code relies on any of these, changing the model ID won't be enough. The first three also apply to Claude Fable 5.1, so code you have already moved to Fable is most of the way there.

Adaptive thinking is always on; Claude effort sets its depth

You can't switch thinking off on Opus 5.5. A request that disables it, or that sets budget_tokens, is rejected with a 400 error. Use adaptive thinking and set the depth with output_config.effort.

Two other changes won't raise errors but will change what you get back. Default effort is medium instead of high, and the model tends to think more per turn than Opus 5 at the same setting. Anthropic's advice is to rerun your effort sweep rather than keep the Opus 5 setting.

Text that Opus 5 wrote between tool calls now arrives inside thinking blocks, which are empty under the default display setting. An interface that showed those progress notes may go quiet after the switch.

Forced tool use and old Claude computer use calls now fail

Setting tool_choice to any or to a specific tool is rejected with a 400 error, while auto and none still work. If your code forces a tool call, that path has to change.

For computer use, the Claude API and Google Cloud now accept only computer_toolset_20260801. The old computer_20251124 tool is refused there. Amazon Bedrock still runs the old tool on Opus 5.5, so Bedrock users don't need to change it.

Claude API rejects thinking blocks after edited history

A thinking block now belongs to the model and the conversation that produced it. If you change the system prompt, the tool definitions or an earlier message and then send old thinking blocks again, the request is rejected with a 400 error.

The check is on by default for accounts created on or after August 31, 2026 (UTC). It applies on the Claude API and on cloud platforms. If your app trims or rewrites earlier turns, look at how it handles thinking blocks before you switch.

Summary

If you use Claude Opus 5.5 in the Claude apps, you can switch now; nothing on your side breaks. If you call it through the API, check the four changes above against your code before you change the model ID. Requests without an effort setting now run at medium, so measure quality and cost again, whether you call Anthropic directly or go through AWS, Google Cloud or Microsoft Azure, where Opus 5.5 is also live.

Sonnet 5.5 and Haiku 5.5 are due in the coming weeks, with no dates announced. For how the current lineup compares, see Claude Models Comparison: Fable, Opus, Sonnet. The previous generation is covered in Claude Opus 5: Near-Fable Output at Half the Price.

Anthropic publishes these announcements and docs as web pages. To save a page as clean text for notes or a prompt, use the tool below.