Claude Opus 5: Anthropic doubles performance at constant price, but lags behind Fable 5

Released on July 24, 2026, Claude Opus 5 maintains the pricing grid of Opus 4.8 ($5 / $25 per million tokens) while doubling the scores on Frontier-Bench v0.1 and achieving Fable 5 on CursorBench 3.2 at half the price. Anthropic's positioning: "frontier intelligence at half the cost."

The concrete case

A tech lead who has already connected their CI pipeline to Claude Sonnet 4.6 or Opus 4.8 wonders, with each Anthropic announcement: "Is it worth changing the ANTHROPIC_MODEL variable?". With Opus 5, released on July 24, 2026, the answer starts to resemble "yes, almost immediately" - provided you look at the right benchmarks.

What changes concretely

Anthropic publishes two joint announcements: the public release note and the technical system card. The guiding principle: "double the performance without touching the price".

Three figures to remember, taken from the official blog post:

• Frontier-Bench v0.1: Opus 5 doubles the score of Opus 4.8, at a lower inference cost.
• ARC-AGI 3: 3× higher score than the second best model in the benchmark - the most impressive metric of the announcement.
• OSWorld 2.0 (agents driving an OS): surpasses Claude Fable 5 at one third of its price.
• CursorBench 3.2 (code assistance in the Cursor editor): in the top 0.5% of Fable 5, at half the price.

The pricing remains identical to Opus 4.8: $5 per million input tokens, $25 per million output. A "fast mode" doubles the throughput (2.5×) against a doubling of the price - useful for Claude Code or interactive pipelines where latency kills the experience.

Where Opus 5 is not first

Reading the announcement honestly, Anthropic acknowledges two limitations:

1. Fable 5 (Anthropic's former internal model released mid-2026) remains ahead of Opus 5 on some benchmarks - Opus 5 approaches it "within 0.5%" on CursorBench, but does not surpass it everywhere.
2. Mythos 5 (the direct competitor, with no further public details) remains above Opus 5 on cybersecurity exploitation tasks. Anthropic does not hide this - it is written in black and white in the announcement, which is notable in a model release communiqué.

For a dev building a code audit agent, this point matters: Opus 5 will be excellent at generation and review, less so on the "I'll find a CVE in this Rust project" part.

Availability

• API: model identifier claude-opus-5.
• Claude.ai, Claude Code, Claude Cowork - all updated on the release day.
• By default on Claude Max subscriptions.
• Best option available for Claude Pro subscribers (above Sonnet).

No information in the public announcement about the size of the context window - a topic that will quickly resurface in dev discussions, especially for ingesting large repos with Claude Code.

Key takeaways

If you are already on Opus 4.8, switching to claude-opus-5 is a no-brainer: same price, doubled performance on agentic benchmarks. If you are on Sonnet for cost reasons, this release does not disrupt your equation - but Opus 5's fast mode is worth testing on interactive tasks where latency matters. And keep in mind that Anthropic itself acknowledges that Opus 5 is not the leader everywhere, especially on the offensive security side.