Anthropic’s Opus 5 Nears Frontier Intelligence at Opus Prices

Anthropic released Claude Opus 5 on July 24, 2026, casting its newest Opus-class model as an everyday workhorse that, by the company’s own account, comes close to the intelligence of its far pricier frontier model, Fable 5, at half the cost. Opus 5 is available immediately across Anthropic’s apps, its API, and the major cloud platforms, at $5 per million input tokens and $25 per million output tokens, the same rates as the model it replaces, Opus 4.8.

The launch makes Opus 5 the default model on Anthropic’s Max subscription and the strongest option available on Pro. The company is pitching it less as a capability ceiling than as a model tuned for daily use, with an effort setting that lets developers trade reasoning depth for lower token cost and faster responses.

What Anthropic’s benchmarks show

On its own coding and knowledge-work evaluations, Anthropic reports that Opus 5 is state-of-the-art on tests including Frontier-Bench and GDPval-AA, while staying behind its restricted Mythos 5 model on cybersecurity tasks. The company says Opus 5 more than doubles Opus 4.8’s score on Frontier-Bench at a lower cost per task, lands within 0.5% of Fable 5’s peak on the CursorBench coding evaluation at maximum effort while costing half as much per task, and triples the next-best model’s result on ARC-AGI 3, a test built around novel problems the model has not seen. On the OSWorld computer-use benchmark, it says Opus 5 beats Fable 5’s best result at just over a third of the cost.

Anthropic frames the jump as a matter of judgment rather than raw scores. It describes a test in which Opus 5 was handed a drawing of a machine part but given no way to view the image; the model wrote its own computer-vision pipeline to pull the geometry from the raw pixels and rebuild the part, a task the company says no competing model solved in five attempts.

Those are vendor-reported figures, run on Anthropic’s own harnesses, and none has yet been independently replicated. That distinction matters more than usual here, because the headline claim is not that Opus 5 tops a leaderboard but that it delivers frontier-adjacent results at a mid-tier price. The one outside check Anthropic discloses is cyber-range testing of early Opus 5 checkpoints by the UK’s AI Security Institute, reported in the model’s system card.

Filling a gap in the lineup

The release plugs a hole Anthropic had opened in its own 2026 lineup. Sonnet 5, out June 30, 2026, delivered close to Opus-level work at a fraction of the price, while the Mythos-class Fable 5, released June 9, 2026, sat above Opus at twice the cost. That left the Opus tier squeezed from both sides. Opus 5 gives Max, Team, and Enterprise customers a frontier-adjacent option without paying Fable or Mythos rates.

For teams running long agentic workloads, cost per completed task tends to matter more than the sticker price per token, and Anthropic is leaning on that math. It says Opus 5 finishes work in fewer steps and with fewer tool calls than Opus 4.8, and it offers a Fast mode that runs about 2.5 times the default speed at twice the base price.

Safety, and the caveats in the system card

Anthropic’s system card calls Opus 5 its most aligned model to date on the company’s automated behavioral audit, scoring ahead of Sonnet 5, Opus 4.8, and the frontier-tier Fable 5 on adherence to its model constitution, with the lowest rate of cooperation with misuse of any model it tested. Those are self-assessments from Anthropic’s internal auditing, not external findings.

The same document is unusually candid about the model’s limits. It notes that Opus 5 “hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall,” and that it trails Mythos 5 on both biology research and offensive cybersecurity. In one internal biology exercise, the model repeatedly got stuck in self-verification loops and failed to deliver a set of protein designs that Mythos 5 completed.

On cyber safeguards, Anthropic modeled Opus 5’s classifiers on Fable 5’s but expects them to intervene about 85% less often. The model is now allowed to hunt for vulnerabilities in source code, work that skews defensive, while scanning of compiled binaries, penetration testing, and exploit generation stay blocked. Requests that do trip a classifier fall back to Opus 4.8 by default. Unlike Fable and Mythos, Opus 5 carries no data-retention requirement for general access.

The practical test now is whether independent evaluators reproduce the coding and reasoning gains Anthropic is reporting, and whether the alignment improvements hold up outside the company’s own audit. Until then, the clearest thing Opus 5 changes is the price of near-frontier capability in Anthropic’s own catalog.