Claude Sonnet 5: API pricing, specs, and agent evaluation
Claude Sonnet 5 targets coding and long-running agents. Review its API changes, temporary pricing, 1M-token context, and production evaluation plan.
Claude Sonnet 5 is Anthropic's June 2026 Sonnet model for coding, tool use, and
long-running agent work. It exposes a one-million-token context window,
adaptive thinking, and the API model ID claude-sonnet-5.
Anthropic positions it close to Opus-class performance on some agentic tasks at a lower price. That claim makes Sonnet 5 worth testing, but it does not remove the need to evaluate the complete application around it.
Claude Sonnet 5 specs at a glance
| Specification | Claude Sonnet 5 |
|---|---|
| Release date | June 30, 2026 |
| API model ID | claude-sonnet-5 |
| Context window | 1,000,000 tokens |
| Maximum synchronous output | 128,000 tokens |
| Inputs | Text and images |
| Thinking | Adaptive by default |
| Effort levels | Low, medium, high, xhigh, max |
| Main workloads | Coding, agents, professional knowledge work |
Introductory API pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. Anthropic says standard pricing becomes $3 input and $15 output after that date. Budget forecasts should also account for prompt caching, batch discounts, tools, and the number of steps in an agent trajectory.
The tokenizer change affects real cost
Anthropic reports that the updated tokenizer can map the same input to roughly 1.0 to 1.35 times as many tokens, depending on the content. A lower headline price can therefore produce a different bill than expected.
Replay representative prompts and record actual input, cached-input, thinking, and output usage. Compare cost per completed task, not a spreadsheet built from one synthetic prompt.
Adaptive thinking changes the tuning surface
Sonnet 5 chooses how much reasoning to use within the configured effort level. Higher effort can improve difficult tasks, but may add latency and token cost. The best setting often varies by route:
| Route | Reasonable starting point |
|---|---|
| Classification or extraction | Low |
| Grounded support answer | Medium |
| Multi-tool workflow | High |
| Hard repository-wide coding task | xhigh, then compare with Opus |
Routing should be evidence-driven. A single global effort setting usually overpays for easy work or underpowers difficult work.
What to test before migrating
Freeze the prompt, tools, retrieval corpus, runtime, and evaluators. Then replay the same production-derived dataset on the old and new model.
Measure:
- End-to-end task completion
- Tool choice, arguments, and ordering
- Groundedness and citation accuracy
- Policy compliance and escalation
- Latency to first useful output and total duration
- Token usage and cost per successful outcome
- Recovery from failed tools or incomplete context
Coding evaluations should include tests, hidden assumptions, brownfield code, and the ability to verify a fix. Customer-facing evaluations should include ambiguous requests, angry users, stale knowledge, and actions that require approval.
Long context is a capability, not a retrieval strategy
A million-token window can hold a large repository or document set, but it can also hide the relevant evidence and increase cost. Compare three modes:
- Full context for tasks that genuinely require global understanding
- Retrieved context for focused questions
- Cached stable context plus newly retrieved evidence
Track which evidence the answer actually used. A correct-looking response that ignored the authoritative source is a future regression waiting to happen.
Monitor Sonnet 5 in production
Capture every model call, tool result, retry, error, and final user outcome. Slice quality by route and effort level. A model can improve coding while regressing on support tone, or improve average quality while increasing tail latency enough to hurt users.
Currai connects production traces to evaluations and alerts so teams can test Sonnet 5 against Sonnet 4.6, Opus, or other providers without losing the exact evidence behind a score. Start with the Currai integration skill.