Anthropic · Model release
Claude Sonnet 5: mid-tier tokens that can hide agentic cost
Sonnet 5 looks cheaper per token than Opus 5, but long agentic runs can produce outsized output bills — the decision is completed-task cost, not the rate card alone.
What changed
- Anthropic positioned Sonnet 5 as the everyday tier with a million-token context window.
- Effort controls and thinking behaviour change latency and spend on harder tasks.
- The model is available through the Anthropic API and cloud catalogues.
Best fit
- Everyday generation
- Short well-scoped tasks
- Teams standardising on Anthropic
What to be careful about
- Recorded DeepSWE cost per task can exceed Opus 5 when retries and output size balloon.
- Per-token savings disappear on long agentic loops.
Dataset snapshot: Claude Sonnet 5 is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.