Claude Opus 5.5: Every Benchmark, From Anthropic, Cursor, Artificial Analysis, Sonar and ARC Prize
Anthropic shipped Opus 5.5 on 22 September 2026 at $4 in and $20 out per million tokens, 20% cheaper than Opus 5. Its own charts say it beats Fable 5.1 almost everywhere. Here is what those charts show, and what Cursor, Artificial Analysis, Sonar and the ARC Prize found when they tested it themselves.
Anthropic released Claude Opus 5.5 on 22 September 2026, three weeks after Fable 5.1. The pitch is unusual for a flagship launch: it is cheaper than the model it replaces, more than 30% faster, and on most of Anthropic's own benchmarks it beats Fable 5.1, a model that costs 150% more per token. Launch charts are marketing, though, so this post puts Anthropic's numbers next to the independent ones that arrived in the first 48 hours.
Anthropic's own numbers
Anthropic compared Opus 5.5 against its own Fable 5.1 and Opus 5 and against OpenAI's GPT-6 Astra and GPT-5.6 Sol, all at maximum effort. Opus 5.5 leads on seven of the nine benchmarks. The two it loses are both to GPT-6 Astra: AutomationBench, by 1.4 points, and Terminal-Bench-Science, by almost six.
- Opus 5.5
- Fable 5.1
- Opus 5
- GPT-6 Astra
- GPT-5.6 Sol
Show the 40 numbers behind this chart
| Category | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — no reading | 41.7% |
| AutomationBench | 40% | 31.4% | 26.9% | 41.4% | 28.8% |
| Humanity's Last Exam | 67.7% | 65.6% | 63.6% | 57.2% | — no reading |
| Terminal-Bench-Science | 58.7% | 52.6% | 29% | 64.6% | 22.4% |
| OSWorld 2.0 | 81.8% | 80.7% | 74% | — no reading | — no reading |
| Chartography | 89% | 88.4% | 83.4% | — no reading | — no reading |
The biggest jump is on Terminal-Bench 4.0, the agentic coding test run in a real shell: 66.4% against 52.3% for Opus 5 and 55.8% for Fable 5.1. That is a 14 point gain in one release. Terminal-Bench-Science doubles, from 29.0% to 58.7%. On the tests that were already near their ceiling, OSWorld 2.0 and Chartography, the gains are a point or less over Fable 5.1.
GDPval-AA, the professional knowledge work benchmark, is scored as an Elo rating rather than a percentage, so it gets its own chart. Here Opus 5.5 is not close to anything: 111 points above Fable 5.1 and more than 300 above GPT-6 Astra.
Show the 5 numbers behind this chart
| Category | Elo |
|---|---|
| Opus 5.5 | 1846 |
| Fable 5.1 | 1735 |
| Opus 5 | 1708 |
| GPT-5.6 Sol | 1588 |
| GPT-6 Astra | 1542 |
Head to head: Opus 5.5 vs Fable 5.1, GPT-6 Astra and GPT-5.6 Sol
The big chart above is dense, so here is each rival on its own, using only the benchmarks where both models have a published score. All numbers are at maximum effort from Anthropic's launch table, which means OpenAI's scores were run on Anthropic's setup. Keep that in mind when a gap is only a point or two.
Opus 5.5 vs Claude Fable 5.1
This is the awkward one for Anthropic, because Fable 5.1 is its own flagship from three weeks ago and costs 2.5 times more. Opus 5.5 beats it on every benchmark in the table. The gaps are large on agentic work (Terminal-Bench 4.0 by 10.6 points, AutomationBench by 8.6) and close to noise on the saturated ones (OSWorld 2.0 by 1.1, Chartography by 0.6).
- Opus 5.5
- Fable 5.1
Show the 16 numbers behind this chart
| Category | Opus 5.5 | Fable 5.1 |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% |
| FrontierCode v1.1 | 54.4% | 50.3% |
| CursorBench 4.0 | 57.8% | 51.8% |
| AutomationBench | 40% | 31.4% |
| Humanity's Last Exam | 67.7% | 65.6% |
| Terminal-Bench-Science | 58.7% | 52.6% |
| OSWorld 2.0 | 81.8% | 80.7% |
| Chartography | 89% | 88.4% |
One caveat: Anthropic evaluated Fable 5.1 with its production safeguards switched on, and said at the Fable launch that this probably costs it some points on agentic tests. Artificial Analysis agrees on the overall ranking, though: 58 for Opus 5.5 against 53 for Fable 5.1.
Opus 5.5 vs GPT-6 Astra
Astra is the only rival that still wins anything. It leads on Terminal-Bench-Science by 5.9 points and on AutomationBench by 1.4, so for long scientific tool use and business process automation it is still the one to beat. Opus 5.5 takes Terminal-Bench 4.0 by 8.5 points, FrontierCode by 1.1 and Humanity's Last Exam by 10.5, and the knowledge work Elo by 304 points.
- Opus 5.5
- GPT-6 Astra
Show the 10 numbers behind this chart
| Category | Opus 5.5 | GPT-6 Astra |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 57.9% |
| FrontierCode v1.1 | 54.4% | 53.3% |
| AutomationBench | 40% | 41.4% |
| Humanity's Last Exam | 67.7% | 57.2% |
| Terminal-Bench-Science | 58.7% | 64.6% |
Two independent numbers soften the Terminal-Bench lead. When Artificial Analysis ran Terminal-Bench 4.0 itself, the two models tied at 59.6%. And Astra reached its scores with about 27,000 output tokens per task against 119,000 for Opus 5.5 at max, so on thinking heavy work its higher per token price is partly paid back by using fewer tokens.
Opus 5.5 vs GPT-5.6 Sol
Sol is the fairest price comparison, because it costs the same per token: $4 in and $20 out (OpenAI's promotional price through November). On that like for like basis, it is not close. Opus 5.5 leads on every shared benchmark, by 29.1 points on Terminal-Bench 4.0, 16.1 on CursorBench and 36.3 on Terminal-Bench-Science.
- Opus 5.5
- GPT-5.6 Sol
Show the 10 numbers behind this chart
| Category | Opus 5.5 | GPT-5.6 Sol |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 37.3% |
| FrontierCode v1.1 | 54.4% | 47.5% |
| CursorBench 4.0 | 57.8% | 41.7% |
| AutomationBench | 40% | 28.8% |
| Terminal-Bench-Science | 58.7% | 22.4% |
Where the lead is, in one chart
This chart shows how many points Opus 5.5 is ahead (above zero) or behind (below zero) each rival on the benchmarks they share. It makes the pattern plain: a small, consistent lead over Fable 5.1, a mixed picture against Astra, and a wide lead over Sol.
- vs Fable 5.1
- vs GPT-6 Astra
- vs GPT-5.6 Sol
Show the 18 numbers behind this chart
| Category | vs Fable 5.1 | vs GPT-6 Astra | vs GPT-5.6 Sol |
|---|---|---|---|
| Terminal-Bench 4.0 | 10.6 pts | 8.5 pts | 29.1 pts |
| FrontierCode v1.1 | 4.1 pts | 1.1 pts | 6.9 pts |
| CursorBench 4.0 | 6 pts | — no reading | 16.1 pts |
| AutomationBench | 8.6 pts | -1.4 pts | 11.2 pts |
| Humanity's Last Exam | 2.1 pts | 10.5 pts | — no reading |
| Terminal-Bench-Science | 6.1 pts | -5.9 pts | 36.3 pts |
What each one costs
Price is where the comparison gets lopsided. Opus 5.5 and Sol share the same token price, while Fable 5.1 and Astra both cost 2.5 times more. Opus 5.5 also has the cheapest cache reads of the four, which matters most on long agent sessions. Note that Sol and Astra both charge extra once a request passes 272K input tokens; the Claude models do not.
- Opus 5.5
- GPT-5.6 Sol
- Fable 5.1
- GPT-6 Astra
Show the 12 numbers behind this chart
| Category | Opus 5.5 | GPT-5.6 Sol | Fable 5.1 | GPT-6 Astra |
|---|---|---|---|---|
| Input | 4 | 4 | 10 | 10 |
| Output | 20 | 20 | 50 | 50 |
| Cache read | 0.2 | 0.4 | 0.25 | 1 |
Cursor: the new top model on CursorBench
Cursor's benchmark matters more than most because it is not built from puzzles. CursorBench 4.0, published on 10 September, is drawn from real production sessions in the Cursor editor: ambiguous, multi file edits, refactors, investigations and design adherence problems. Cursor announced Opus 5.5 as its new number one at 57.8% on Max, and said it costs 40% less per task than Opus 5.
Show the 4 numbers behind this chart
| Category | Score |
|---|---|
| Opus 5.5 | 57.8% |
| Fable 5.1 | 51.8% |
| Opus 5 | 46.6% |
| GPT-5.6 Sol | 41.7% |
The 40% saving per task is bigger than the 20% cut in the per token price, and the difference is efficiency: Opus 5.5 finishes the same tasks in fewer tokens. That lines up with what Sonar measured below.
Artificial Analysis: first place by five points
Artificial Analysis runs its own suite of ten evaluations and rolls them into one Intelligence Index. Opus 5.5 at max effort scores 58, the highest on the board, with GPT-6 Astra and Fable 5.1 tied at 53 and Opus 5 at 51. It leads on six of the ten evaluations, including Humanity's Last Exam at 61.4%, SciCode at 66.9% and an Elo of 1822 on AA-Briefcase, against 1679 for Fable 5.1.
Show the 4 numbers behind this chart
| Category | Index score |
|---|---|
| Opus 5.5 | 58 |
| GPT-6 Astra | 53 |
| Fable 5.1 | 53 |
| Opus 5 | 51 |
The more useful view is what each effort setting buys. Artificial Analysis tested all five, and the curve is steep at the bottom and flat at the top. Pick an effort level below to see the score, the cost per index task and the output speed.
Going from low to medium adds nine index points for 79 cents more per task. Going from xhigh to max adds two points for $2.52 more. For most production work, medium or high is where the value is, and medium at 51 already matches Opus 5 at its best.
There is one number in the Artificial Analysis data that cuts the other way. At max effort, Opus 5.5 used about 119,000 output tokens per task, against 78,000 for Fable 5.1, 73,000 for Opus 5 and only 27,000 for GPT-6 Astra. At max, it thinks a lot.
Show the 4 numbers behind this chart
| Category | Output tokens |
|---|---|
| Opus 5.5 | 119k |
| Fable 5.1 | 78k |
| Opus 5 | 73k |
| GPT-6 Astra | 27k |
Why the independent scores are lower
Put Anthropic's numbers next to Artificial Analysis and the same benchmarks come out lower when someone else runs them. Terminal-Bench 4.0 is 66.4% in Anthropic's table and 59.6% in Artificial Analysis's run, where GPT-6 Astra ties it at 59.6%. Humanity's Last Exam is 67.7% for Anthropic (with tools) and 61.4% for Artificial Analysis. Different harnesses, tool access and time limits explain most of that. It does not change who comes first, but on agentic coding the lead over GPT-6 Astra is much smaller than the launch chart suggests.
- Anthropic
- Artificial Analysis
Show the 4 numbers behind this chart
| Category | Anthropic | Artificial Analysis |
|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 59.6% |
| Humanity's Last Exam | 67.7% | 61.4% |
Sonar: less code, fewer tokens, slightly more bugs per line
Sonar ran Opus 5.5 through 544 HumanEval and MBPP tasks with executable tests and then scanned every line it wrote. The pass rate is basically unchanged: 87.7% against 88.6% for Opus 5. What changed is how it gets there. Opus 5.5 wrote 27.5% less code and used 40% fewer output tokens to do it.
Show the 6 numbers behind this chart
| Category | Change vs Opus 5 |
|---|---|
| Output tokens | -40% |
| Lines of code | -27.5% |
| Bugs (total) | -19% |
| Vulnerabilities (total) | -34% |
| Code smells (total) | -43% |
| Bugs per line | 11.8% |
Read that chart carefully. Opus 5.5 produces 428 bugs where Opus 5 produced 528, and 152 vulnerabilities where Opus 5 produced 230. But because it writes so much less code, the bugs it does write are packed more densely: 644 per million lines against 576. Vulnerability density goes the other way, down 9%, and code smells per thousand lines drop 21%. Cognitive complexity per line rises 4%. The fair summary is tighter, terser code that you still need to review.
ARC Prize: past 90% on ARC-AGI-2
The ARC Prize verified Opus 5.5 on its semi private sets at every effort level. The headline is 93.3% on ARC-AGI-2 and 98.5% on ARC-AGI-1, both at high effort. Notably, high beats both xhigh and max on each set, a sign that on these puzzles more thinking stops helping past a point. No ARC-AGI-3 score was published.
- ARC-AGI-1
- ARC-AGI-2
Show the 10 numbers behind this chart
| Effort | ARC-AGI-1 | ARC-AGI-2 |
|---|---|---|
| Low | 88.5% | 70.1% |
| Medium | 97.5% | 87.5% |
| High | 98.5% | 93.3% |
| Xhigh | 97.5% | 92.5% |
| Max | 97.5% | 91.7% |
Price: cheaper per token, cheaper per task
Every line on the price sheet came down. Input and output are 20% cheaper than Opus 5, cache writes 20% cheaper, and cache reads drop 60%, from $0.50 to $0.20 per million. Fast mode costs double: $8 in and $40 out. Anthropic's own estimate is that the lower price and the higher speed together make typical workloads about 40% cheaper than on Opus 5, which matches Cursor's per task figure.
- Opus 5.5
- Opus 5
Show the 8 numbers behind this chart
| Category | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | 4 | 5 |
| Output | 20 | 25 |
| Cache write | 5 | 6.25 |
| Cache read | 0.2 | 0.5 |
The comparison that will get people talking is Fable 5.1, which costs $10 in and $50 out. Opus 5.5 is 60% cheaper per token and scores higher on most of the benchmarks above. Unless you need something only Fable does, it is hard to argue for paying 2.5 times more.
What else is new
- Five effort levels: low, medium (the default), high, xhigh and max.
- Better prompt injection resistance and better persistence on long, multi step tasks.
- Clearer written output, which Anthropic calls out as a deliberate change in how the model communicates.
- The same cybersecurity and biology safeguards and anti distillation protections introduced with Fable 5.1, plus zero data retention for eligible customers.
- The EU AI Act watermark that every model released after 2 August 2026 carries.
The verdict
Every independent test so far points the same way as Anthropic's charts: Opus 5.5 is the strongest generally available model right now, and it is cheaper than what it replaces. Cursor puts it first on real editor work, Artificial Analysis puts it first overall, and the ARC Prize puts it above 90% on ARC-AGI-2. The caveats are real but small. Its lead on agentic coding shrinks when someone else runs the test, it writes slightly buggier code per line, and at max effort it spends a lot of tokens. Start at medium or high, measure, and only go to max when the task earns it.
Sources: Anthropic, Cursor, Artificial Analysis, Sonar, ARC Prize, Vellum, OfficeChai and OpenAI's GPT-5.6 Sol pricing. For the earlier head to head between the two September flagships, see GPT-6 Astra and Claude Fable 5.1: The Benchmarks Side by Side.
