benchr
Claude Opus 5 arrived on July 24, 2026. On paper, the move from Opus 4.8 looks unusually simple. The list price and headline limits stay familiar. One request-setting change is where a careless migration can fail.
Replay the request before you flip the alias
The model ID is claude-opus-5. Adaptive thinking is on by default, and Anthropic says a request that tries to disable it at xhigh or max effort returns a 400 error. Save a real request from production and replay it in staging. That check will tell you more about migration risk than another broad capability chart.
Same price does not mean same bill
The list rate is $5 input and $25 output per million tokens. The 1M-token context window and 128K output ceiling leave room for long coding and research jobs, but they do not tell you how many tokens a successful job will consume. Price the prompts and completions you already run, then compare a small, scored batch with your current Claude deployment.
There is no benchmark table to hide behind
Anthropic calls Opus 5 a step change over Opus 4.8, but the source record does not include a benchmark table benchr can reproduce. Keep the choice grounded in your own pass/fail cases, latency, token totals, and operational failures. A provider description alone does not earn the model a rank.
| Field | Verified record |
|---|---|
| API model ID | claude-opus-5 |
| Input / output list price | $5 / $25 per 1M tokens |
| Context / maximum output | 1M / 128K tokens |
| Adaptive thinking | On by default; some disablement requests return 400 |