The reported outcome
Anthropic introduced Claude Opus 5.5 on September 22. It says an early tester completed a 680,000-line migration in under a day, a striking example of the scale coding agents are being asked to handle. The claim comes from a tester cited by Anthropic; the announcement does not make that one project a universal benchmark.
Anthropic also reports that, when tasked with reducing load times across a web app, Opus 5.5 succeeded in 39 of 40 attempts. It presents the model as a step up from Opus 5 for complex, long-running work, with performance near Claude Fable 5.1 on many tasks.
Cost and control
Anthropic says a typical workload costs 40% less to run than Opus 5. Published API prices are $4 per million input tokens and $20 per million output tokens, each 20% below Opus 5. Cache reads cost $0.20 per million tokens. Anthropic also reports output generation more than 30% faster than its predecessor.
For a migration, cost per token is only one line of the bill. Engineers still need tests, code review, deployment checks and rollback plans. The material advance is that a capable model may sustain a broader change while producing work that remains inspectable. A fast migration that subtly changes behavior would be expensive to fix.
What to watch
Anthropic says external evaluators including METR tested Opus 5.5 before release, and that the model improved on its own behavioral audit. These are useful signals, but product evaluation should include the specific repository, build and review process in which the model will be used.
The most revealing comparison is a real backlog task: let the old and new models attempt the same migration, then measure working code, review time, regressions and total expense. If the reported large-scale result repeats across teams, agents will become more useful for maintenance work that has long been postponed because it is too tedious to do manually.
Source published 2026-09-22. Coverage is based on the maker’s announcement and demonstration.