Today, Anthropic released Claude Opus 5. The flagship Opus-tier product, it replaces Claude Opus 4.8. Prices remain unchanged at $25 per million tokens for output tokens and $5 per million tokens.
Anthropic teams position Opus 5 in a way that it is similar to Claude Fable 5 but at half price. This is the default version of Claude Max, and the most powerful model in Claude Pro.
What actually changed at the API level
Three changes are quite important before any benchmark does:
- Thinking is on by default. On Opus 4.8, requests ran without thinking unless you set
thinking: {"type": "adaptive"}. On Opus 5 the same request thinks, and the effort parameter controls depth. Becausemax_tokenscaps thinking plus response text, existing values need review. - There is a breaking change. Setting
thinking: {"type": "disabled"}with effortxhighormaxnow returns a 400 error. This restriction applies per request. You either cap effort athighor drop thethinkingfield. - Anthropic tells developers to delete their verification prompts. Instructions like “include a final verification step” now cause over-verification, because the model already verifies its own work. The Opus 5 prompting guide covers the tuning patterns.
The model ID is claude-opus-5. The default context is 1M tokens, and there are no other variants. The maximum output for the Synchronous Messages API is 128k Tokens. The Message Batches API reaches 300k with the output-300k-2026-03-24 beta header. The minimum caching prompt has been reduced to 512 tokens from 1,024.
Coding and agentic results
On FrontierBench v0.1, a 74-task successor to Terminal-Bench 2.1, Opus 5 scored 43.3% at max effort. Opus 4.8 achieved 18.7%. Fable 5 scored 33.7%, while GPT-5.6 sol reached 37.5%. At xhigh effort Opus 5 reaches 44.4% mean reward, its best result.
A detail is noteworthy. Opus 5 safety classifications refused and flagged 5% API calls across 4 trials. Fable classifiers marked 42% of all calls over 26% trials.
Opus 5 scored 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro. Fable 5 has a slight edge on Pro, at 80%. Multimodal SWE bench has seen a bigger jump from 38.4% to 59%.
These agentic results are where the biggest wins can be seen. Opus 5 reached 70.57% on OSWorld 2.0 against 55.7% for Opus 4.8. On Zapier AutomationBench it scored 26.0%, against 17.0% for Opus 4.8 and 17.4% for Fable 5. Even at medium effort, it scores 24% for $0.89 per task.
On Artificial Analysis GDPval-AA v2, Opus 5 takes the top two leaderboard spots at ELO 1861 and 1827. The xhigh setting beats every other model while using 25% fewer output tokens than max.
Reasoning, and the ARC-AGI-3 result
Anthropic prompted Opus 5 on all six IMO 2026 problems without tools or an agent harness. All 24 solutions generated by a three-model panel were scored correct. Experts independently graded each pre-specified problem solution at 7/7. Above the cutoff of 29/42, 42/42 represents gold medal level.
According to the ARC Prize Foundation, a score of 30.16 % on ARC AGI-3 was verified at high effort. This is about four times higher than the previous best leaderboard score. GPT 5.6 Sol achieved 7.78% while Opus 4.8 attained 1.52%. The results of Opus 5’s maximum effort are not available at the release.
On Humanity’s Last Exam, Opus 5 scored 56.3% without tools and 64.7% with them.
Multimodal section contains a valuable lesson. The cost-effectiveness of using agentic tooling to scale test-time computing is greater than that achieved by adaptive thinking.
Opus 5’s score on Chartography was 29.6% with no tools but 83.0% using an image-cropping and container tool. On BenchCAD Vision2Code, voxel IoU moves from 0.366 to 0.821. This is a huge improvement over Claude Mythos 5, which was 0.678.
‘;
for(i=0;i

