Anthropic quietly upgraded its lineup this week with Opus 5 — a faster, cheaper workhorse that, in many tests, bests the company’s own flagship, Claude Fable 5.
Why it matters
- Opus 5 arrives as a practical alternative to Fable 5, which has been beset by availability and regulatory headaches since its June debut. Fable 5 was pulled days after launch amid a U.S. emergency export-control order over a jailbreak vulnerability, and when it returned it was removed from standard subscriptions and moved to a credits-only model.
- Opus 5 slots into Anthropic’s four-tier lineup underneath the restricted “Mythos” class (which houses Fable 5 publicly and a tighter, vetted Claude Mythos 5 through Project Glasswing). Anthropic positions Opus as the heavy-duty option between the cheaper Haiku and mid-tier Sonnet models.
Benchmarks: Opus 5 leads on most fronts
Independent and partner tests show Opus 5 outperforming Fable 5 and other commercial rivals on multiple robust benchmarks:
- Frontier-Bench v0.1 (end-to-end software engineering tasks): Opus 5 — 43.3% passed; Fable 5 — 33.7%; OpenAI’s GPT-5.6 Sol — 34.4%.
- ARC-AGI-3 (novel problem-solving puzzles): Opus 5 — 30.2%; GPT-5.6 Sol — 7.8%. (Fable 5 was not tested here.)
- GDPval-AA v2 (knowledge-work Elo rating): Opus 5 — 1,861; Fable 5 — 1,747; GPT-5.6 Sol — 1,736.
- AutomationBench (Zapier workflow automation): Zapier reports Opus 5 completed end-to-end churn-prevention workflows — identifying at-risk accounts, notifying owners, and summarizing for retention teams — delivering a 100% pass where prior models failed.
Developer and research feedback
- Lovable, a major developer platform, found Opus 5 not only stronger on hard, agentic coding tasks (Fabian Hedin reported a 22% gain over Opus 4.7) but also far more consistent across runs.
- Ultima Genomics praised Opus 5’s behavior in long-form scientific analysis: better statistical thinking, cross-checking results, and sticking to multi-step workflows.
A few caveats
- Fable 5 still holds small advantages in narrow legal and health tasks, where it outperforms Opus by a slim margin.
- The competing open-weight model Kimi K3 from Moonshot AI (2.8 trillion parameters) landed a week earlier; independent benchmarks generally place Kimi K3 behind Fable 5 and GPT-5.6 Sol overall, though it beats them in specific areas.
Pricing and availability
- Opus 5 is available now via API: $5 per million input tokens and $25 per million output tokens. A Fast mode (~2.5× speed) is also offered at double price: $10/$50 per million tokens.
- Unlike Fable 5’s restricted Mythos positioning, Opus 5 is accessible via subscription for broader customer use — potentially easing concerns over the public availability of Anthropic’s higher-capability models.
Bottom line
Opus 5 looks designed to be the practical frontier model Anthropic needed: cheaper to run, broadly available, and demonstrably stronger on many real-world engineering and automation benchmarks — even as Fable 5 retains narrow edge cases in legal and health domains.
Read more AI-generated news on: undefined/news
