Claude Opus 5 lands at half of Fable 5's price and beats it on eight of thirteen benchmarks

Anthropic shipped Claude Opus 5 on July 24, two months after Opus 4.8 landed on May 28, and it arrives priced at roughly half of Fable 5 while beating it on 8 of 13 internal benchmarks. The interesting number isn't the win rate. It's the cadence: a full Opus generation released in the time it used to take to publish a system card.

The direct comparison against Fable 5 is where Anthropic wants readers to look. On Frontier-Bench, the 74-task suite that stretches across physics, chemistry, and cryptography, Opus 5 scores 9.7% higher. Anthropic also claims state-of-the-art results on GDPval-AA, and reports that Opus 5 outperformed Opus 4.8 across every internal life-sciences benchmark, with automated protein research called out by name. Mythos 5, the internal cybersecurity benchmark Anthropic still declines to release publicly on cybersecurity grounds, remains ahead of Opus 5 on cyber tasks.

Opus 5 also scores zero on an internal Anthropic benchmark that evaluates a model's ability to exploit vulnerabilities. Anthropic frames that as a deliberate ceiling, not a capability gap, and the shape of the guardrails backs it up: Opus 5 will search source code for defensive vulnerability review, but refuses to scan compiled binaries. Cyber gains, the company says, come as a side effect of general capability rather than targeted training.

The commercial pitch is the effort dial. Operators can set low, medium, or high per request, trading intelligence for tokens, and Anthropic publishes performance curves against it. This is the same posture OpenAI took with GPT-5.6 on July 9: economy as a first-class API parameter, not a marketing footnote.

The second shipped feature is quieter and, for anyone who ran production traffic through Fable 5, more consequential. Automatic Fallbacks lets a new fallbacks parameter with a default mode reroute prompts that Opus 5's safety classifier refuses, sending them to a less-restricted model instead of returning an error. Anthropic expects those classifiers to engage 85% less often than they did on Fable 5. That's the vendor conceding, in numbers, that Fable 5's refusal rate was a budget problem as much as a safety one. Fable 5's token consumption pushed customers through budgets fast enough that Anthropic pulled it on June 12 and re-released it on June 30 after security work.

Opus 5's safeguards match Opus 4.8 in most areas and are stronger on cyber. Anthropic describes it as its safest model to date, less prone to deceptive behavior and to hard-to-reverse errors than its predecessors. Mid-conversation tool changes, adding or removing tools between turns while preserving the prompt cache, enter beta, the first time the Opus tier has offered it.

Opus 5 becomes the new default on Claude Max and the top option on Claude Pro. What Anthropic has actually shipped is a pricing correction with a safety story attached.

Sources

• https://www.anthropic.com/news/claude-opus-5
• https://techcrunch.com/2026/07/24/anthropic-launches-opus-5/
• https://www.bloomberg.com/news/articles/2026-07-24/anthropic-unveils-more-cost-efficient-model-for-everyday-tasks
• https://fortune.com/2026/07/24/anthropic-debuts-claude-opus-5-with-feature-that-lets-users-toggle-between-cost-and-capability/
• https://siliconangle.com/2026/07/24/anthropic-launches-claude-opus-5-efficiency-safety-improvements/