Claude Opus 5 is live, and it arrives with a clear proposition: near-frontier intelligence at roughly half the cost of Claude Fable 5. Priced identically to its predecessor Opus 4.8 at $5 per million input tokens and $25 per million output tokens, Opus 5 becomes the new default on Claude Max and the strongest option on Claude Pro. What sets it apart is not raw benchmark bragging rights alone, but a shift in how the model behaves: more deliberate, more thorough, and noticeably better at verifying its own work before handing it back.

What Claude Opus 5 actually changes

The headline story is efficiency paired with capability. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s score at a lower cost per task. On CursorBench 3.2 at max effort, it lands within 0.5% of Fable 5’s peak score at half the cost per task. That combination, top-tier reasoning without top-tier spend, is what makes Opus 5 a practical daily driver rather than a specialist tool reserved for the hardest jobs.

The effort setting matters here. Developers can dial reasoning up when a task warrants it or dial it down to conserve tokens for faster, cheaper output. Even at its lowest effort setting on Zapier’s AutomationBench, Claude Opus 5 passes more business tasks end to end than any competing model. At its ceiling, it delivers roughly 1.5× the pass rate of the next-best model for the same cost.

Standout benchmark results

  • ARC-AGI 3: Opus 5 scores three times higher than the next-best model on novel problem solving.
  • OSWorld 2.0: outperforms every other model at any given cost on computer-use tasks, beating Fable 5’s best result at roughly a third of the cost.
  • Life sciences evaluations: gains across structural biology, organic chemistry, and bioinformatics, with a 10.2 point jump on molecular structure inference from spectroscopy data and 7.7 points on protein variation analysis.

How Opus 5 performs in production workflows

Feedback from early-access customers points to a common theme: the biggest gains show up on longer, vaguer, more open-ended work. On genomics analysis, Opus 5 reaches for the right statistical tests to rule out confounders and cross-checks its own results using independent methods. On financial modeling, it averaged 9 percentage points higher accuracy across effort levels while using a third fewer turns and tool calls, cutting total time by 60%.

Frontend and full-stack development see similar improvements. On the same app builds, Opus 5 produces the cleanest animations, games, and 3D work of any Opus model. It also catches its own frontend bugs by inspecting rendered pages at desktop and phone widths, spotting hidden elements and off-screen buttons before returning the work.

Legal teams testing Opus 5 on agent work saw the largest gains in corporate governance and arbitration, alongside a 26% reduction in tokens generated compared to Opus 4.8 at max reasoning. Box reported an 8% overall improvement over Opus 4.8 on enterprise content analysis, with data analysis workflows up 11% and due diligence up 17%.

Alignment and safety posture

Anthropic’s automated behavioral audit rates Opus 5 as its most aligned model to date. It adheres to Claude’s Constitution more consistently than Opus 4.8, Sonnet 5, or Fable 5, shows the lowest rates of deceptive behavior across the family, and is the least susceptible to being tricked into misuse. It also demonstrates the strongest track record so far on avoiding reckless actions with hard-to-reverse consequences.

On dual-use capabilities, Opus 5 remains behind Mythos 5 in both biology research and offensive cybersecurity. Anthropic has intentionally avoided training Opus 5 on cyber tasks, though the model’s general capability gains mean it now approaches Mythos 5 at finding vulnerabilities. It stays substantially behind on exploit development, the step that turns a vulnerability into an actual threat.

Guardrails and fallbacks

Cyber classifiers on Opus 5 are proportionally less restrictive than those on Fable 5, expected to intervene around 85% less often. They allow source-code vulnerability discovery but block binary-based scanning, penetration testing, and exploit generation. Flagged requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default, and API users can enable the same fallback behavior. For biology-related requests blocked on Fable 5, traffic now routes to Opus 5 rather than Opus 4.8, making Opus 5 the most capable generally available model for scientific research.

What ships alongside the model

Two beta updates arrive with Opus 5. Mid-conversation tool changes on the Claude Platform let developers swap available tools without invalidating the prompt cache, a small change that removes a real friction point for agent workflows. Automatic fallbacks on the API give users the option to route safety-flagged requests to another model instead of receiving a block, keeping pipelines running on the best available model rather than stalling.

Opus 5 is also offered in Fast mode, running roughly 2.5 times the default speed at twice the base price on the Claude Platform, or through usage credits in Claude Code. For teams already running Opus 4.8, migration should be straightforward: the pricing is unchanged, and the model ID is claude-opus-5 on the API.