Anthropic announces Claude Opus 5.5 with claimed cost and performance gains
On September 22, 2026, Anthropic introduced Claude Opus 5.5, described as the first model in its new Claude 5.5 family and positioned as a successor to Opus 5 with performance closer to Claude Fable 5.1 on many tasks. Anthropic says Opus 5.5 was externally evaluated by Frontier Design and METR and is the company’s first release following its public call to pace the frontier. [1]
Performance and cost‑effectiveness
Anthropic reports that Opus 5.5 is a substantial step up from Opus 5, claiming large downstream gains on complex work: early testers reportedly completed multi‑day engineering tasks in hours, and the company highlights examples such as a 680,000‑line code migration completed in under a day by one tester. Anthropic also states that Opus 5.5 performs well at finding and fixing software inefficiencies and that it scored higher than other Claude models in an exercise building a game from a single prompt. [1]
The company emphasizes efficiency: Opus 5.5 is said to use less compute than Opus 5 and, at default settings, to cost about 40% less on typical workloads while generating output more than 30% faster. Anthropic provides per‑million token pricing and comparative charts showing Opus 5.5 matching or exceeding other models on agentic coding benchmarks at a fraction of the cost per task. [1]
| Price type | Opus 5.5 (per 1M tokens) | Opus 5 (for comparison) |
|---|---:|---:|
| Cache reads | $0.20 | $0.50 |
| Input tokens | $4 | $5 |
| Output tokens | $20 | $25 |
| Cache writes | $5 | $6.25 |
| Fast mode (input/output) | $8 / $40 | n/a |
(Prices reported by Anthropic.) [1]
Safety, audits, and verification
Anthropic reports that Opus 5.5 achieved the strongest results to date on its automated behavioral audit, an alignment suite run across thousands of simulated scenarios, and claims the model is less likely than recent models to take hard‑to‑reverse actions, act outside its bounds, or succumb to prompt injection. The company says it broadened its alignment testing to include longer and impossible tasks and scenarios modeled on real incidents, and points readers to the Opus 5.5 System Card for full evaluation details. [1]
Because Anthropic characterizes Opus 5.5 as comparable to Mythos 5.1 in biology and cybersecurity domains, it is deploying Opus 5.5 with safeguards similar to those used for Fable 5.1 and is gating sensitive uses through verification programs. Vetted organizations can apply now to a Life Sciences Verification Program, and Anthropic says it will expand its Cyber Verification Program to verified cybersecurity practitioners in coming weeks. [1]
Developer impact and practical considerations
Anthropic frames Opus 5.5 as particularly strong for long, sprawling coding jobs, citing examples where Opus 5.5 completed audits and migrations faster and with fewer tokens than Opus 5 and Fable 5.1. In one internal comparison rewriting HAProxy from C to Rust, Anthropic says both rewrites passed nearly all regression tests but that Opus 5.5 finished faster and at lower cost than Fable 5.1. If these reported efficiencies are realized in production, teams running large agentic coding or persistent developer agents could see materially lower per‑task costs and shorter runtimes. [1]
Anthropic also says it is increasing usage limits on several paid plans and offering a subscription rate‑limit reset to help users take advantage of Opus 5.5’s lower cost and faster generation. These changes may affect budgeting and throughput for teams using Claude in continuous integration, automated testing, or other agentic pipelines. [1]
Developers evaluating Opus 5.5 should note Anthropic’s caveat that benchmark margins are an imperfect guide to real‑world differences at these capability levels and that some benchmark runs were affected by safeguard interventions; Anthropic reports that when safeguards intervened on certain biology or cybersecurity evaluations, lower‑capability models completed the tasks, which can reduce measured performance. [1]