Claude Opus 5 Benchmarks: Wins, Losses, and Where Opus 5.5 Leads

Claude Opus 5 scores 51 on the Artificial Analysis Intelligence Index (v4.3.2) in max effort mode, which puts it #11 of 211 models and one point ahead of Claude Fable 5 at 50. Anthropic released the model on July 24, 2026, and backed it with wins on four named benchmarks, plus one loss to Mythos 5 on cybersecurity. Claude Opus 5.5, released September 22, 2026, now leads the same index at 58.

Checked September 26, 2026 against Artificial Analysis and Anthropic's model pages.

That loss matters as much as the wins. Anthropic's own launch post places Opus 5 behind Mythos 5 on cybersecurity tasks, the one benchmark category where the company did not claim the top spot. Here is the full rundown, benchmark by benchmark, with the source for every number.

Claude Opus 5 benchmark results at a glance

Artificial Analysis Intelligence Index: Opus 5 beats Fable 5, Opus 5.5 now leads

The Artificial Analysis Intelligence Index combines 10 evaluations into one comparative score. Version 4.3.2, the current one, uses AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA, Terminal-Bench 4.0, SciCode, Humanity's Last Exam, GDP.pdf, CritPt, AA-Omniscience and AA-LCR v1.1. Version 4.3 replaced the earlier banking evaluation with AutomationBench-AA and moved to Terminal-Bench 4.0, so scores from older index versions do not line up with today's numbers.

On v4.3.2, Opus 5 in its max effort configuration scores 51 and ranks #11 of 211 models. Fable 5 (max effort) scores 50 and ranks #13. So Opus 5 still edges out Fable 5 by one point. It is no longer at the top: Claude Opus 5.5 (max effort) scores 58 and ranks #1, and Artificial Analysis now flags Opus 5 as superseded by it. An earlier version of this article reported Opus 5 at 61 and #1 of 170. Those figures are out of date.

A one-point gap over Fable 5 is not a landslide. What makes it notable is the price attached to it: Opus 5 lists at $5 input and $25 output per million tokens, half of Fable 5's $10 and $50, so the higher scorer of the two is also the cheaper one. The price side is covered in more depth in our Claude Opus 5 pricing breakdown.

Frontier-Bench: more than double Opus 4.8

Anthropic says that on Frontier-Bench v0.1, Opus 5 surpasses all other models and more than doubles the performance of its prior flagship, Opus 4.8, at a lower cost per task. Anthropic has not published the full underlying task list for Frontier-Bench in the same detail as third-party suites like Artificial Analysis, so read this as a self-reported comparison, a useful data point, not an independently verified leaderboard result.

ARC-AGI 3: 3x the next-best model

ARC-AGI is designed to test reasoning that resists memorization: each task asks the model to solve novel problems it has not seen in training. Anthropic reports Opus 5's score on ARC-AGI 3 at three times the next-best model's.

This is the widest claimed margin in the entire release. A 3x gap on a benchmark built to resist pattern-matching from training data is a stronger signal than a similar margin on a benchmark closer to typical training data, though as with Frontier-Bench, this is Anthropic's own reported comparison rather than an independently published leaderboard.

OSWorld 2.0: computer use at a third of the cost

OSWorld 2.0 measures how well a model can operate a real computer environment: navigating applications, completing multi-step UI tasks, and recovering from errors without step-by-step human instructions. Anthropic says Opus 5 outperforms every other model at any given cost on this benchmark, surpassing Fable 5's best result at just over a third of the cost.

This is the clearest bridge between the benchmark numbers and the pricing story. Anthropic states the comparison as cost, not as the sticker price per million tokens, which matters for computer-use work where the model takes many sequential actions per task. For the token-price side of that comparison, see our Opus 5 pricing guide.

Zapier AutomationBench: 1.5x the next-best model

Zapier's AutomationBench measures whether models can complete business tasks from start to finish, the kind of multi-step, tool-calling work that runs inside no-code and low-code automation platforms. Anthropic reports Opus 5's pass rate at around 1.5x the next-best model's for the same cost per task.

Of the five benchmarks in this release, AutomationBench is the most directly relevant to teams building or running AI agents inside automation tools, since it is scored on tasks that resemble production automation work rather than academic reasoning problems. Artificial Analysis now includes an AutomationBench-AA run inside its own index.

Where Opus 5 does not win: cybersecurity

Anthropic's launch post is specific about this one: Opus 5 remains behind Mythos 5 on cybersecurity tasks. On OSS-Fuzz, one of Anthropic's cybersecurity evaluations, Opus 5 is close to Mythos 5 at identifying software vulnerabilities but far behind at developing exploits for them. Anthropic also says it intentionally avoided training Opus 5 on cyber tasks, as it did with Opus 4.8.

The same post adds, outside the benchmark results, that Mythos 5 remains the stronger model for biology research. If your workload is security research, vulnerability analysis, or red-team tooling, this is the category in the launch where Opus 5 is not the strongest choice inside Anthropic's own lineup. Mythos 5 is offered separately, by invitation only, for defensive cybersecurity work.

How Opus 5 compares to Fable 5 overall

Putting the results together: on the one third-party index that names both models (the Artificial Analysis Intelligence Index), Opus 5 comes out one point ahead, and on the one benchmark in Anthropic's Opus 5 launch post with a direct Fable 5 comparison (OSWorld 2.0), Opus 5 wins on both capability and cost. Anthropic's launch post places Opus 5 slightly behind Fable 5's peak on CursorBench 3.2, and its Fable 5.1 table shows Opus 5 below Fable 5 on HLE (no tools) and CursorBench 3.2.0. The cybersecurity loss is against Mythos 5, a different model. And Opus 5 does this at half Fable 5's token price.

For the full side-by-side of the two models, capability and price together, see our Claude Opus 5 vs Claude Fable 5 comparison. If you are also weighing non-Anthropic options, our Claude Fable 5 alternatives roundup covers models outside this family, and our models page has current specs and pricing for the full Anthropic lineup.

Update: Claude Opus 5.5 versus Opus 5 (Anthropic's figures)

Anthropic released Claude Opus 5.5 on September 22, 2026, and now lists Opus 5 as a legacy model that is still available. Every number in this section comes from Anthropic's Opus 5.5 launch page and its model documentation. We have not run these tests ourselves.

Anthropic notes that its Terminal-Bench 4.0 result for Opus 5.5 was run at xhigh effort, and that the default effort on the API dropped from high on Opus 5 to medium on Opus 5.5. On cybersecurity, Anthropic describes Opus 5.5 as comparable to Claude Mythos 5.1 in biology and cybersecurity and ships it with safeguards similar to Fable 5.1's. Moving code over is not only a model ID swap: thinking can no longer be disabled and forced tool use returns an error. Our Claude Opus 5.5 release breakdown covers the price change, the breaking changes, and whether to migrate.

A note on how to read self-reported benchmarks

Four of the five Opus 5 results in this rundown (Frontier-Bench, ARC-AGI 3, OSWorld 2.0 and Zapier AutomationBench) come from Anthropic's own launch materials rather than an independently maintained public leaderboard. That does not make them false, but it means a third party has not reproduced them the way Artificial Analysis runs its own index. The Intelligence Index score (51 on v4.3.2, #11 of 211) is the one figure here you can check yourself on Artificial Analysis's model page, independent of Anthropic. It also moves: index versions change and new models arrive, which is exactly why the score in this article changed after Opus 5.5 shipped.

FAQ

What is Claude Opus 5's Intelligence Index score?

Claude Opus 5 scores 51 in max effort mode on the Artificial Analysis Intelligence Index v4.3.2, ranking #11 out of 211 models, as checked September 26, 2026. Claude Fable 5 scores 50, one point behind. Claude Opus 5.5 leads the index at 58.

Is Claude Opus 5 better than Claude Fable 5?

On the Artificial Analysis Intelligence Index v4.3.2, Opus 5 scores 51 against Fable 5's 50, a narrow win, and Anthropic reports Opus 5 beating Fable 5's best OSWorld 2.0 result at just over a third of the cost. Its other launch results (Frontier-Bench, ARC-AGI 3, Zapier AutomationBench) are framed against "the next-best model" or "all other models" rather than Fable 5 by name, and cybersecurity is the area where Anthropic places Opus 5 behind Mythos 5, not Fable 5.

Does Claude Opus 5 beat Mythos 5?

Not across the board. Anthropic's July 24, 2026 launch post places Opus 5 behind Mythos 5 on cybersecurity tasks: close at finding vulnerabilities on OSS-Fuzz, far behind at developing exploits. The same post says Mythos 5 remains the stronger model for biology research.

What is ARC-AGI 3 and why does Opus 5's score matter?

ARC-AGI 3 is a reasoning benchmark in which the model has to solve novel problems it cannot have memorized from training data. Anthropic reports Opus 5 scoring three times the next-best model on it, the widest margin in this release, which is a stronger signal of reasoning gains than similar margins on more familiar benchmark formats.

How good is Claude Opus 5 at computer use?

Anthropic reports Opus 5 surpassing Fable 5's best OSWorld 2.0 result at just over a third of the cost. On Anthropic's Opus 5.5 launch page, Opus 5 scores 74.0% (partial) on OSWorld 2.0 against 81.8% for Opus 5.5. See our Opus 5 pricing guide for the token-price side.

Are Anthropic's Opus 5 benchmark claims independently verified?

Partially. The Artificial Analysis Intelligence Index score (51 on v4.3.2, #11 of 211) comes from a third-party leaderboard you can check directly. Frontier-Bench, ARC-AGI 3, OSWorld 2.0 and Zapier AutomationBench results are Anthropic's own self-reported figures from its launch materials, not yet reproduced on an independent public leaderboard.

Is Claude Opus 5 good for AI agents and automation?

Anthropic reports Opus 5's pass rate at around 1.5x the next-best model on Zapier's AutomationBench, which measures end-to-end business tasks, plus a computer-use win on OSWorld 2.0 at lower cost. Both point toward agentic and automation workloads. For new builds, Anthropic's own table shows Opus 5.5 at 40.0% on AutomationBench against Opus 5's 26.9%.

Was Claude Opus 5 the only model Anthropic released recently?

No. Anthropic introduced Claude Fable 5 and Claude Mythos 5 together, with Fable 5 released June 9, 2026, then Opus 5 on July 24, 2026. Claude Opus 5.5 followed on September 22, 2026, and Anthropic now lists Fable 5.1 and Opus 5.5 as its current Fable and Opus models.

Is Claude Opus 5 still the top model on the Intelligence Index?

No. As of September 26, 2026, Claude Opus 5.5 (max effort) leads the Artificial Analysis Intelligence Index v4.3.2 at 58, and its xhigh and high effort settings hold the next two spots. Claude Opus 5 (max effort) scores 51 and sits at #11 of 211 models on the same index version.

How much better is Claude Opus 5.5 than Opus 5?

On Anthropic's own launch table, Opus 5.5 beats Opus 5 on all nine listed benchmarks, including Terminal-Bench 4.0 (66.4% vs 52.3%), AutomationBench (40.0% vs 26.9%) and OSWorld 2.0 (81.8% vs 74.0%). It also costs less: $4 input and $20 output per million tokens against Opus 5's $5 and $25. See our Opus 5.5 release breakdown.

Should I still use Claude Opus 5 now that Opus 5.5 is out?

Opus 5 is still available as a legacy model, with retirement not sooner than July 24, 2027, so existing integrations keep working. Anthropic recommends migrating to Opus 5.5, which is cheaper per token, but check the breaking changes first: thinking can no longer be disabled, and forced tool use returns an error.

---

Sources: Anthropic, Introducing Claude Opus 5, Anthropic, Claude Opus 5.5, Claude Opus 5 model page, Claude Opus 5.5 model page, What's new in Claude Opus 5.5, Artificial Analysis, Claude Opus 5, Artificial Analysis, Claude Opus 5.5, Artificial Analysis LLM leaderboard, CNBC, Anthropic's Claude Opus 5 launch, TechCrunch, Anthropic launches Opus 5