Claude Opus 5.5
Opus 5.5 shipped on 22 September 2026 and took Opus 5 place in Anthropic own table, at $4 in and $20 out per million tokens. It shows a large jump on the maker own benchmarks, but it has not yet received an independent Arena vote.
- 40% cheaper than Opus 5
- no Arena vote yet
- default effort dropped from high to medium
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. What changed
Where it moved against the previous version
- Terminal-Bench 4.0 52.3%66.4%
- CursorBench 4.0 46.6%57.8%
Both figures were published by the maker itself, with no independent leaderboard behind them. Whether the move is good is what the table and the text below say, not the length of these segments.
The price that actually moved this time
From Opus 4.5 through Opus 5, five straight versions sat at the same $5 and $25. Opus 5.5 breaks that: $4 in, $20 out, forty percent cheaper by Anthropic own account. Cache reads fell too, from fifty cents to twenty cents, a quarter of the old price.
Anthropic also says output now generates more than thirty percent faster. For anyone costing a service on real consumption, that means a smaller invoice and a faster answer, not just a new number on the price list.
A jump Anthropic actually measured
Unlike Opus 5, which shipped only relative claims, this launch carries absolute figures, and Anthropic printed Opus 5 own score on the same tests alongside them:
| Test | Opus 5.5 | Opus 5 |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 52.3% |
| CursorBench 4.0 | 57.8% | 46.6% |
| FrontierCode v1.1 | 54.4% | 48.0% |
| GDPval-AA v2.1 | 1846 | 1708 |
| AutomationBench | 40.0% | 26.9% |
The largest relative jump is AutomationBench, 26.9 to 40.0, roughly fifty percent better. The smallest is FrontierCode, 6.4 points. The pattern is familiar: the longer and more multi-step the task, the bigger the gap.
One caveat worth keeping: every one of these figures was measured and published by Anthropic itself. No independent leaderboard backs them yet.
What still does not show up on our ranking table
This reference ranks purely from the independent Arena leaderboards, never from a maker own launch numbers. As of today, 23 September 2026, neither the web-dev board nor the text board carries any row under this model name. So Opus 5.5 does not appear in the coding or chat tables yet, strong self-reported benchmarks or not. The moment Arena rates it, it enters.
One less-obvious point for anyone with code already tuned on Opus 5: the default effort dropped from high to medium. Code that never set effort explicitly now behaves differently with no error raised, and needs re-testing.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
- API model ID
- claude-opus-5-5 Source
- Released
- 22 September 2026 Source
- Context window
- 1 million tokens Source
- Max output
- 128k tokens Source
- Input price
- $4 per million tokens Source
- Output price
- $20 per million tokens Source
- Cached input price
- $0.20 per million tokens Source
- Knowledge cutoff
- June 2026 Source
- Open weights
- no Source
- Input
- text, image Source
- Output
- text Source
- Where it ships
- Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS Source
Using it from Iran
This column carries a date and says how it was checked, because most listicles guess it. Where we have only read a vendor policy page, the note below says exactly that.
- Reachable
- blocked
- Payment
- no working route
- Free tier
- no
- Measured on
How we checked: Iran is not on Anthropic supported-countries list. Read from Anthropic own page rather than an independent network test. Nothing changed for this version.
What it is good at
- A real jump on Anthropic own agentic benchmarks, such as Terminal-Bench 4.0 from 52.3 to 66.4 percent
- Cheaper than Opus 5: $4 and $20 instead of $5 and $25, with cache reads at a quarter of the old price
- Shipped on five routes, including Claude Platform on AWS, where Opus 5 carried a dash
- The newest knowledge cutoff of any Anthropic model, tied with Fable 5.1: June 2026
Where it falls short
- No independent Arena vote yet, so it does not appear in our own ranking table and every figure above is the maker own measurement.
- Default effort dropped from high to medium; code written for Opus 5 that never set effort explicitly now behaves differently, with no error.
- Disabling thinking no longer works; a request that sets it returns a 400 error.
- No official route from Iran, same as the rest of the Opus family.
Our take
On the numbers Anthropic itself published, this version should take Opus 5 place in our coding table too. But this reference only moves on independent evidence, never a maker claim, so it stays here for now: announced, not yet ranked.
Questions people actually ask
Does Opus 5.5 replace Opus 5
Yes. In Anthropic own table Opus 5 now counts as a legacy version and Opus 5.5 is current. Opus 5 is still reachable on every route, it simply is not the latest any more.
Why has it not moved your ranking
Because our ranking comes only from the independent Arena leaderboards, never from a maker announcement. Until Arena rates this model, it does not appear in the coding or chat tables, however strong the self-reported benchmarks are.
Did the price go up or down
Down. Input fell from $5 to $4 and output from $25 to $20, and cache reads dropped from fifty cents to twenty cents.