Agent Capability Now Has a Runtime Dial
Claude Sonnet 5’s important signal is not one more benchmark lead. Its effort control lets developers steer token use and tool activity within one model—making runtime policy part of agent capability, while stopping short of a guaranteed compute budget.
Anthropic released Claude Sonnet 5 on June 30 across Claude plans, Claude Code and its API. The API launched at an introductory $2 per million input tokens and $10 per million output tokens through August 31, before moving to $3/$15; official model documentation lists a 1-million-token context window and 128,000-token maximum synchronous output.13
The interesting part is not the sticker price. Sonnet 5 exposes five effort levels—low, medium, high, xhigh and max. Anthropic says this setting changes all response tokens, including thinking and tool calls; lower effort tends to make fewer tool calls, while higher effort can explore more. It is explicitly a behavioral signal, not a strict token budget.4
| Evidence | Sonnet 5 | Sonnet 4.6 | What to remember |
|---|---|---|---|
| BrowseComp, single agent | 84.7% | 76.2% | Sonnet 5 used max effort, a 10M-token limit and context compaction |
| OSWorld-Verified | 81.2% | 78.5% | 361 tasks, 100 steps, pass@1 averaged over five runs |
| SWE-bench Pro | 63.2% | 58.1% | Vendor-run; standard configuration used adaptive thinking at max effort |
| API list price after Aug. 31 | $3/$15 per MTok | $3/$15 per MTok | Sonnet 5 may tokenize the same input to roughly 1.0–1.35× the previous token count |
The table supports a narrower conclusion than “cheap model beats expensive model.” Anthropic’s own system card says most capability evaluations used 300,000 to 1 million tokens, while BrowseComp used a 10-million-token ceiling with compaction.2 BrowseComp itself contains 1,266 deliberately hard-to-find web questions; its authors describe it as incomplete but useful, and note that it avoids real-user complications such as ambiguity and long-form answers.5 OSWorld is more operational, but Anthropic also changed its harness by fixing a zoom bug and increasing the per-turn token limit before re-evaluating the older model.26
So the defensible insight is about runtime policy. Choosing the checkpoint no longer fixes how an agent will behave: the same model can make fewer or more tool calls and spend fewer or more response tokens depending on effort. A production system can start routine work at low or medium effort, then escalate selected tasks when its own checks show that the extra cost improves results.
The engineering question is therefore measurable: which tasks improve enough at each effort level to justify the extra tokens, tool calls and latency? Anthropic’s documentation also warns that changing effort within a cached conversation invalidates cached prefixes. Teams should sweep effort levels on their own evaluations and hold the setting constant where prompt-cache economics matter.4