Anthropic's Opus 5.5 Beats Fable on Benchmarks — Output Tokens Now $20/M
First Claude Opus release since Dario Amodei backed pacing the frontier. Faster, less jargon; Sonnet 5.5 and Haiku 5.5 still weeks out. Bio/cyber safeguards match Fable/Mythos — company claims, not independent bake-offs.
Anthropic released Claude Opus 5.5 on Tuesday, claiming state-of-the-art results on coding and knowledge-work benchmarks — and, on many of those tests, a lead over the larger Fable model. Primary reporting: TechCrunch (Russell Brandom).
What shipped (claimed)
- Opus 5.5 positioned as SOTA for coding and knowledge work
- Outpaces larger Fable on many benchmarks; also on informal tasks Fable failed
- Output tokens cut to $20/M from $25 on prior Opus; other metrics also cheaper
- Faster to serve / less compute to run than prior Opus
- Messaging: less jargon, important information first
- About two months after Opus 5 (July 24); Sonnet 5.5 and Haiku 5.5 still weeks out
- Bio/cyber risk profile described as comparable to Mythos → same safeguards as Fable
- First Claude Opus release since CEO Dario Amodei publicly embraced pacing the frontier
Caveat: Performance and safety claims are company claims. External eval partners named (METR, Frontier Design) — treat as company-sponsored, not independent bake-offs until third-party results land.
Price and ops signal
- Output: $20 per million tokens (was $25)
- Other usage metrics also lower
- Serve cost / latency claimed down versus prior Opus
For operators buying frontier coding/knowledge capacity, the headline is less "new god-model" and more: same family, lower output unit cost, faster serving, clearer prompts — while mid-tier Sonnet/Haiku refreshes are still pending.
Operator takeaway
If your stack already runs Opus for coding or knowledge workflows, re-check unit economics on output-heavy jobs before you expand seat count. Re-run a short bake on your own eval set — vendor benches are not your task bundle. Keep leave / augment / automate on the work itself: which reviews, drafts, and lookups actually move when jargon drops and latency falls. Soft next step if useful: an Ada conversation to map which Opus-dependent tasks should migrate first once Sonnet 5.5 lands.
Source: TechCrunch — Anthropic releases Opus 5.5 with lower prices and Fable-level performance (Russell Brandom, Sep 22, 2026).
