Anthropic announced Claude Opus 5 on 24 July 2026. In the company's own framing it is "a step change improvement for the Opus tier powering long-running agents while delivering improvements in coding and professional work."
Worth noting given the week it arrived into: Opus 5 had been rumoured for a 20–21 July launch, then reported as missed, before landing on the 24th. If you read a post last week saying it did not exist, that post was right when it was written.
The interesting word is "long-running"
Anthropic leads with long-running agents rather than a benchmark number, and that is a meaningful choice of emphasis. The hard problem in agent work has stopped being whether a model can do a task and started being whether it can stay coherent across hundreds of steps without drifting, looping, or quietly losing the plot.
A model that is marginally smarter per step but materially better at step 200 is a different product from one that wins a single-turn evaluation. That is the pitch here.
What we are not going to tell you yet
We have not run our own testing against Opus 5, so we are not publishing a score for it or repeating vendor benchmark figures as though we had verified them. Our 361 score for Claude stays at 9.1 until we have put the new model through the same process as everything else on this site.
Be sceptical of anyone publishing a confident hands-on verdict within hours of a model launch. There has not been time.
Who should care immediately
- Teams running agents that execute for minutes or hours rather than seconds — this release is explicitly aimed at you.
- Anyone whose Claude workload is long-document analysis, where the Opus tier already justified its cost.
- Cursor and Claude Code users, since coding improvements are called out directly in the announcement.
Who can wait
If your usage is short chat turns, drafting, or summarising, the Opus tier was probably already more model than the task needed. A cheaper tier will keep being the right answer, and Opus 5 does not change that arithmetic.