Meet the New Claude Opus 5: Frontier-Class Agentic Coding and Computing at Fixed Opus Prices

Today, Anthropic released Claude Opus 5. It replaces Claude Opus 4.8 as the flagship of the Opus-tier. The price is fixed at $5 for one million input tokens and $25 for one million output tokens.

The Anthropic team pitches the Opus 5 as approaching the brilliance of the Claude Fable 5 at half the price. It is now the default model in Claude Max and a powerful model in Claude Pro.

What exactly has changed at the API level

Three changes are very important before any benchmark can be performed:

1. Thinking is on by default. In Opus 4.8, applications run without thinking without you setting thinking: {"type": "adaptive"}. In Opus 5 the same request is considered, and the effort parameter controls the depth. Because max_tokens thinking caps and response text, existing values ​​need to be updated.
2. There is a breaking change. Setting up thinking: {"type": "disabled"} with effort xhigh or max now it returns a 400 error. The limit is applied on a per request basis. You or cap effort at high or discard thinking field.
3. Anthropic tells developers that remove their confirmation commands. Commands such as “insert final validation step” now cause over-validation, because the model is already validating its own function. Opus 5’s inspiring guide includes tuning patterns.

The model ID is claude-opus-5. Core is 1M tokens as both default and maximum, with little variation. The maximum output is 128k tokens for the synchronous messaging API. Message Batches API up to 300k with output-300k-2026-03-24 beta header. The minimum cacheable prompt drops to 512 tokens, down from 1,024.

Agent code and results

In FrontierBench v0.1, which replaced 74 tasks in Terminal-Bench 2.1, Opus 5 scored 43.3% with maximum effort. Opus 4.8 got 18.7%. Legend 5 reached 33.7% and GPT-5.6 Sol reached 37.5%. In xhigh effort Opus 5 reaches 44.4% means reward, its best result.

One detail from that run is worth noting. Opus 5 security classes flagged and rejected 5% of API calls, out of all 4% attempts. Myth 5 classifiers flagged 42% of calls across 26% of trials.

Opus 5 scored 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro. Legend 5 still puts it in Pro at 80.0%. For SWE-bench Multimodal the jump is even bigger, from 38.4% to 59.4%.

Agent numbers are the clear winners. Opus 5 scored 70.57% in OSWorld 2.0 compared to 55.7% for Opus 4.8. In Zapier AutomationBench it scored 26.0%, compared to 17.0% for Opus 4.8 and 17.4% for Fable 5. In the average effort it still scored 24% at $0.89 per task.

In the Performance Analysis GDPval-AA v2, Opus 5 takes the top two places of the leaderboard at ELO 1861 and 1827. xhigh optimization outperforms all other models while using 25% fewer output tokens. max.

Consultation, and result of ARC-AGI-3

Anthropic caused Opus 5 in all six problems of IMO 2026 without tools or agent harness. A panel of three model judges scored all 24 generated solutions correctly. Human experts independently rated one pre-specified solution for each problem on 7/7. A final 42/42 is the gold medal standard, above the 29/42 cutoff.

The ARC Prize Foundation reports 30.16% guaranteed in the ARC-AGI-3 with great effort. That’s nearly four times the leaderboard result previously reported. GPT-5.6 Sol reached 7.78% and Opus 4.8 reached 1.52%. The results of Opus 5 with a lot of effort were not available when it was released.

In the Ultimate Human Test, the Opus 5 scored 56.3% without tools and 64.7% with them.

The multimodal class has an active course. The Agent’s tool to use scales when testing includes cost-effectiveness rather than dynamic thinking alone.

In Chartography, Opus 5 scored 29.6% without tools and 83.0% with container and image cropping tool. In BenchCAD Vision2Code, the voxel IoU goes from 0.366 to 0.821. In tools, that beats Claude Mythos 5 at 0.678 by a wide margin.

‘; for (i=0; i‘+f[i][1]+’‘+f[i][2]+’