Opus 5.5 leads the benchmarks, but the 40% cost cut is the real release note

Anthropic released Claude Opus 5.5 on September 22, the first model in a new Claude 5.5 family, and framed it as performing "at the level of Claude Fable 5.1 on most work" while costing 40% less to run than Opus 5 on typical workloads. The published pricing: $4 per million input tokens and $20 per million output tokens, 20% below Opus 5, with cache reads — which Anthropic says make up the majority of agentic and coding costs — at $0.20 per million, a 60% cut. Output generation is more than 30% faster. Five-hour usage limits rise on Pro, Max, Team and seat-based Enterprise plans, and subscribers get a bankable rate-limit reset. Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks.

On benchmarks, Anthropic reports leads in agentic coding and knowledge work: 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1 and 52.3% for Opus 5; 57.8% on CursorBench 4.0 versus 51.8%; 1846 on GDPval-AA v2.1 versus 1735. But the same table shows OpenAI's GPT-6 Astra ahead on two axes — 41.4% versus 40.0% on AutomationBench, and 64.6% versus 58.7% on Terminal-Bench-Science. Anthropic itself adds the caveat we would have written for it: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences," and says the gap to Fable 5.1 feels narrower in practice than the scores imply. That is a vendor telling you not to over-read its own scoreboard, and it deserves to be quoted more than the scores themselves.

The most interesting disclosure is buried in a footnote. Opus 5.5 was evaluated with production safeguards enabled, and when those safeguards intervened, cybersecurity tasks were completed by Claude Opus 4.8 and biology and frontier-LLM-development tasks by Opus 5 — which, Anthropic concedes, likely lowered the reported numbers. In other words, the published figure describes a governed system, not a raw model. That is the honest way to benchmark a deployed product, and it should become the norm.

On safety, Anthropic says Opus 5.5 posts its best-ever scores on its automated behavioral audit, is less likely than recent models to take hard-to-reverse actions or step outside granted boundaries, and is more resistant to prompt injection — the attack where instructions hidden in data hijack an agent. External evaluators including Frontier Design and METR tested it pre-release. Because it is comparable to Claude Mythos 5.1 in biology and cybersecurity, it ships behind Fable 5.1-grade safeguards, with vetted organizations applying to a Life Sciences Verification Program and a Cyber Verification Program expanding soon. Anthropic also notes this is its first release since it publicly called for "pacing the frontier" — a tension it does not resolve, and neither should we pretend it is resolved.

Our read: the headline capability claims — one tester finishing a 680,000-line code migration in under a day, a 39-of-40 success rate on cutting page load times — are the labor story in miniature, and we will not soften it. Work that occupied engineering teams for weeks is collapsing into a day, and the adjustment for people whose job was that work is real and near-term. But the mechanism underneath is the one that eventually pays for the optimism: capability rose while cost per unit of work fell 40% in a single generation. Deflation like that is how frontier-grade reasoning stops being a luxury for well-funded labs and becomes infrastructure for a hospital, a school district or a two-person company. The detail we would flag to other labs is the least glamorous one: Anthropic says Opus 5.5 writes more clearly and puts key information first, and it explicitly calls that a safety benefit because its work is easier to follow and check. Legibility as an auditing property, not a style preference, is the kind of quiet design choice that compounds as these systems take on more consequential work.