⚡ AI Focus Bulletin
ModelsIndustrySafety

Anthropic launches Opus 5 at near-Fable 5 performance

Anthropic's Opus 5 approaches Fable 5-class results and triples the field on ARC-AGI-3, all at unchanged Opus pricing and lower token cost.

What launched

Anthropic released Claude Opus 5 on July 24, positioning it as a frontier-adjacent model that matches or approaches Claude Fable 5 across most public benchmarks while costing the same as its predecessor, Opus 4.8 — $5 per million input tokens and $25 per million output tokens, roughly half Fable 5's rate. The model is available immediately on all paid Claude plans and the API, and is the default on Claude Max.

The numbers

The headline result is ARC-AGI-3, a test of novel problem-solving where Anthropic reports Opus 5 scored about three times the next-best system (roughly 30% versus single digits for rival frontier models). On agentic terminal coding via Frontier-Bench, Opus 5 leads Fable 5 and GPT-5.6 by wide margins, though on some execution-based coding suites it lands marginally behind both. Third-party harnesses corroborated the pitch on release day: Cursor measured it within half a point of Fable 5 on CursorBench at half the price, Cognition's Devin put it near Fable-level on FrontierCode 1.1, and Perplexity ranked it second only to Fable 5 on its WANDR agentic benchmark.

Efficiency and safety framing

Anthropic and outside reviewers framed the release as a token-efficiency story rather than a raw capability leap — Ars Technica noted legal workloads used around a quarter fewer tokens than Opus 4.8 at maximum reasoning. Anthropic also called Opus 5 its most aligned model by internal behavioral audit, with the lowest measured rates of reckless behavior, while noting it is stronger than Opus 4.8 on cybersecurity tasks yet still well behind the restricted Mythos 5 on exploit development.

Why it matters

With Fable 5 constrained by US export controls, Opus 5 gives developers a widely available model at close-to-frontier quality for far less money, intensifying the price-per-intelligence race just as OpenAI and Anthropic reportedly prepare for IPOs.

Sources