Anthropic made Claude Opus 5 the default model behind Claude Max subscriptions. It also became the strongest option on Claude Pro, replacing Opus 4.8 across both consumer tiers. Pricing stayed flat at $5 per million tokens in and $25 per million tokens out on the API, matching the prior version exactly. The bet is straightforward. Anthropic wants a cheaper, faster model to handle most of what its flagship, Claude Fable 5, does, so Fable 5 stays reserved for the hardest jobs while Opus 5 absorbs everyday coding and analysis work.
Anthropic says Opus 5 leads its internal Frontier-Bench and GDPval-AA evaluations. Both benchmarks were built and are run by Anthropic itself. On CursorBench 3.2, the company reports Opus 5 landing within half a percentage point of Fable 5’s top score, at roughly half the per-task price. Those are Anthropic’s own numbers, produced on Anthropic’s own test harness. No outside audit has confirmed the margins.
The company also cites established third-party-style tests. On ARC-AGI 3, a benchmark built around unfamiliar problem types, Anthropic says Opus 5’s score triples the next-best model it tested. On Zapier’s AutomationBench, which checks whether a model can carry a business task from intake to resolution, Anthropic puts Opus 5’s completion rate at about 1.5 times the runner-up, at equal cost. On the computer-use benchmark OSWorld 2.0, Anthropic reports Opus 5 beating every rival regardless of price. It even clears Fable 5’s own top mark while costing roughly a third as much to run.
Cybersecurity is the one place Anthropic does not claim the lead. Its release places Opus 5 behind a separate, specialized model called Mythos 5 on offensive security work. On an internal exploit test called OSS-Fuzz, Opus 5 finds vulnerabilities nearly as well as Mythos 5. It lags badly, though, at turning those findings into working exploits. Anthropic frames the gap as deliberate. The company says it kept cyber-specific material largely out of Opus 5’s training, and it expects those classifiers to trigger about 85 percent less than the ones guarding Fable 5.
On its own pre-deployment audit, Anthropic ranked Opus 5 as the least prone to deceptive or reckless behavior among its recent releases. It assigned the model a score of 2.3 on a misalignment scale, where lower is safer. That audit, like the performance numbers, is an internal Anthropic instrument. No outside safety lab has published a matching evaluation yet.
Anthropic’s announcement carries endorsements from coding and workflow partners including Cursor, Devin, Zapier, Box, Lovable, and JetBrains, each describing gains on tasks specific to their own products. Those are testimonials solicited for a launch post, not independently reviewed comparisons. None of the partners disclosed a testing methodology.
Teams currently paying Opus-tier API rates for Opus 4.8 have a concrete reason to re-run their own workloads against Opus 5 before their next contract renewal: Anthropic is charging the same price for what it says is roughly double the coding throughput. Security teams relying on Claude for vulnerability research should read the Mythos 5 gap literally. Opus 5 gets easier to use. Its capacity to turn a found flaw into a working exploit has not measurably improved.
Anthropic published this announcement on its website, Anthropic.com, on July 27, 2026.
