Claude Opus 5

Overview

Claude Opus 5 is Anthropic's flagship model, released on July 24, 2026 and served as claude-opus-5. Anthropic frames it as a step change rather than an incremental release over Opus 4.8: frontier intelligence approaching Claude Fable 5, at the same price as Opus 4.8 and half the price of Fable 5.

The gains concentrate in three areas: deep reasoning, long-horizon agentic coding, and test-time compute scaling. Opus 5 converts extra thinking budget into better results more reliably than any earlier Opus, so the effort level you pick carries more weight than it used to. Thinking is also on by default — no thinking field is needed, and the model decides how much to think each turn.

Key capabilities

Effort matters more. Start at the default high and move in either direction based on your evals: step down where quality holds to save tokens and latency, step up only for the most demanding work. At xhigh or max, raise max_tokens so the model has room to think across subagents and tool calls.

Disabling thinking has a precondition. thinking: {"type": "disabled"} is accepted only at effort high or below; combining it with xhigh or max returns a 400. This is a breaking change from Opus 4.8.

Caching and billing rules. The minimum cacheable prompt drops from 1,024 to 512 tokens. Cache reads bill at 0.1x base input, 5-minute writes at 1.25x, 1-hour writes at 2x, and the Batch API is 50% off both directions. The full 1M window bills at the standard rate with no long-context tier. Fast mode (research preview, Claude API only) is 2x the standard rate.

Note that Claude 4.7 and later use a newer tokenizer that produces roughly 30% more tokens for the same text — re-estimate costs accordingly when migrating. See live pricing in the model catalog.

Benchmarks

Opus 5's benchmark story is agentic: terminal coding, computer use, search, and business-process automation come first, with pure knowledge reasoning second. Across Anthropic's four-model comparison it takes the top score on Frontier-Bench, GDPval-AA v2, ARC-AGI-3, BrowseComp, OSWorld 2.0, AutomationBench, and BioMysteryBench. It trails in only three places, all narrowly or to a specialist: DeepSWE (3.9 points behind GPT-5.6 Sol), FrontierCode and HLE-without-tools (0.1–0.2 points behind Fable 5), and HealthBench (behind the health-specialized Mythos 5).

The chart above is Anthropic's published comparison, with Opus 5 outlined in orange against Fable 5, Opus 4.8, and GPT-5.6 Sol. The ARC-AGI-3 row stands out: 30.2% against 1.5% for Opus 4.8 and 7.8% for GPT-5.6 Sol — an order-of-magnitude jump on genuinely unseen problems rather than an incremental gain.

Agentic coding and terminal work

Frontier-Bench v0.1 exercises the full terminal-agent loop on the mini-SWE-agent harness. Opus 5's 43.3% is more than double Opus 4.8 (21.1%) and ahead of both Fable 5 (33.7%) and GPT-5.6 Sol (34.4%). On real repositories, FrontierCode v1.1 lands at 53.4%, effectively level with Fable 5 (53.5%) at half the cost, while DeepSWE v1.1 reaches 68.8%, behind GPT-5.6 Sol's 72.7%.

This cost-versus-score curve says more than any single number. The x-axis is USD per attempt on a log scale, and the points along each line are the five effort levels from low to max. Opus 5's entire curve sits above Opus 4.8 and Fable 5 — the same spend buys a higher score, or the same score costs less. The curve peaks around xhigh and does not climb further at max, which is exactly why Anthropic recommends starting at high and tuning from evals.

Computer use and business automation

OSWorld 2.0 covers long desktop and browser workflows, where Opus 5 scores 70.6% — nearly 15 points above Opus 4.8. Anthropic notes it passes Fable 5's best result at just over a third of the cost. AutomationBench shows an even wider gap: a 26.0% pass rate, roughly 1.5x the next-best model (GPT-5.6 Sol at 18.1%) for the same cost per task.

Reasoning and knowledge work

GDPval-AA v2 scores deliverable-heavy knowledge work — reports, contracts, research synthesis — in Elo, and Opus 5's 1861 leads Fable 5 (1747) and GPT-5.6 Sol (1736). Humanity's Last Exam with tools is also the highest at 64.7%, 6.8 points above Opus 4.8. On internal domain evals, Anthropic reports Opus 5 scoring 10.2 points higher than Opus 4.8 on organic chemistry, 7.7 points higher on protein tasks, and 9 points more accurate on financial modeling while taking 60% less time.

Alignment and safety

Anthropic calls Opus 5 its most aligned model to date. In an automated behavioral audit (lower is better), Opus 5 scores 2.30, below Opus 4.8 (2.85), Mythos 5 (2.81), and Sonnet 5 (3.35). On the security side, its OSS-Fuzz vulnerability identification rate rises to 79.4% (Opus 4.8: 61.5%), but it fully exploits only 4 challenges versus 13 for Mythos 5, and its cyber classifiers intervene about 85% less often than Fable 5's.

A few behavioral differences are worth planning for: default responses and written deliverables run longer, the model narrates progress more often in agentic sessions, it delegates to subagents more readily in multi-agent frameworks, and it verifies its own work unprompted — carried-over instructions like "include a final verification step" or "use a subagent to verify" cause over-verification and should be removed when migrating.

When to use it

• Long-horizon agentic coding: multi-file feature work, large refactors, and end-to-end fixes — the doubled Frontier-Bench score comes from staying on task through long tool loops without leaving stubs.
• Computer use and process automation: leading scores on both OSWorld 2.0 and AutomationBench make it a strong driver for browsers, desktops, and internal enterprise systems.
• Code review and bug finding: Anthropic highlights a high real-bug rate per pass with few false positives, holding up at lower effort levels — a good fit for always-on CI review.
• Cost-sensitive frontier work: when you need something close to Fable 5's intelligence on half the budget, this is the direct swap.
• Very long context: the 1M window keeps instruction following and tool calling consistent throughout, with no long-context surcharge.
• Multi-agent orchestration: reliable writer-verifier patterns with few cases of agents overwriting each other's work.

CrossModel exposes Claude Opus 5 through Anthropic-compatible /v1/messages and OpenAI-compatible /v1/chat/completions. Current pricing is available in the model catalog.