Claude Sonnet 5 is Anthropic's mid-tier model in the Claude 5 lineup, and for most teams it is the model to default to, not the one you settle for after Opus 5 turns out too expensive. It replaces Sonnet 4.6 at the same list price — $3 per million input tokens, $15 per million output tokens — with an introductory rate of $2/$10 running through August 31, 2026. The bigger change is behavioral: Sonnet 5 runs adaptive thinking by default, supports the full five-level effort dial through xhigh, and closes most of the gap to Opus 5 on coding and agentic work. For day-to-day production traffic — code generation, structured extraction, tool-calling agents, RAG synthesis — Sonnet 5 is now the practical default, and Opus 5 becomes a deliberate upgrade for the hardest tasks rather than a security blanket.
What changed in 2026
- Adaptive thinking is on by default. Sonnet 4.6 shipped with thinking off unless you asked for it; Sonnet 5 reasons through multi-step problems automatically, and you control depth with the effort parameter instead of a manual token budget.
- A full effort ladder, including xhigh. Sonnet 5 is the first Sonnet-tier model with low, medium, high, xhigh, and max effort levels. Anthropic recommends xhigh for the hardest coding and agentic work and high as the default elsewhere.
- A new tokenizer. Sonnet 5 shares its tokenizer with the current Opus generation, and the same text tends to tokenize into noticeably more tokens than it did on Sonnet 4.6. Re-check your max output ceilings and cost dashboards rather than assuming the old numbers carry over.
- Higher-resolution vision. Sonnet 5 is the first Sonnet-tier model with support for higher-resolution images, which matters for screenshot-heavy, chart-heavy, and document-heavy workloads.
- Non-default sampling parameters are rejected. Temperature, top_p, and top_k no longer accept custom values on requests — behavior is steered through prompting instead.
Sonnet 5 at a glance
| Spec |
Claude Sonnet 5 |
| Model ID |
claude-sonnet-5 |
| Context window |
1,000,000 tokens |
| Max output |
128,000 tokens |
| Input price |
$3 / million tokens ($2 introductory through Aug 31, 2026) |
| Output price |
$15 / million tokens ($10 introductory) |
| Thinking |
Adaptive, on by default |
| Effort levels |
low, medium, high, xhigh, max |
The context window and output ceiling are identical to Opus 5's — 1M tokens in, up to 128K out. That matters more than it sounds: in the Sonnet 4.6 era, teams sometimes paid Opus prices purely to get more context headroom. That reason has effectively disappeared with Sonnet 5.
Who should actually use it
Sonnet 5 is the right default for coding assistants, agent loops that run a few dozen tool calls, structured data extraction, summarization, and most chat-tier products that need real reasoning rather than a single quick answer. Anthropic itself is unusually candid that low and medium effort perform well above what their names suggest on this model — for routine work, that's where the real cost savings live, not in avoiding the model tier altogether.
Where Sonnet 5 is not the right call: the hardest, longest-horizon agentic runs such as repo-scale refactors or deep multi-document synthesis, workloads that specifically need Fable 5's always-on maximal reasoning, or extremely high-volume, latency-critical classification where Haiku 4.5 is both cheaper and faster.
Common mistakes
- Assuming thinking is off unless you ask for it. That was true on Sonnet 4.6. On Sonnet 5, omitting the thinking parameter now runs adaptive thinking automatically, which changes both output and cost on requests that used to be thinking-free.
- Reusing Sonnet 4.6 token budgets. The new tokenizer means the same prompt can consume noticeably more tokens. A max output ceiling tuned for 4.6 can truncate output that used to fit comfortably.
- Defaulting to high or xhigh everywhere. It's tempting to treat higher effort as strictly better. On routine tasks, low or medium frequently matches what older models needed high effort to do — test before committing to the more expensive setting.
- Passing temperature out of habit. Requests that set temperature, top_p, or top_k to a non-default value are rejected outright. Strip them and steer style through the prompt instead.
FAQ
Is Claude Sonnet 5 a bigger jump than Sonnet 4.6 was over 4.5?
On coding and agentic benchmarks, yes — Anthropic positions the coding and agentic gains as the headline of this release, closing much of the distance to Opus-tier quality at Sonnet pricing.
Does the introductory pricing apply automatically?
Yes, the discounted $2/$10 rate applies to claude-sonnet-5 traffic through August 31, 2026, with no separate opt-in required, after which standard $3/$15 pricing takes over.
Can I still use manual thinking budgets?
No. The fixed token-budget approach to extended thinking is removed on Sonnet 5; adaptive thinking paired with the effort parameter is the only supported control.
Is Sonnet 5 available everywhere Sonnet 4.6 was?
It launched on the first-party Claude API plus major cloud platforms, but a few betas, such as fast mode, remain Claude-API-only — check current availability rather than assuming full feature parity everywhere.
Where to go next
If you're weighing whether to pay more for the flagship tier, see our full Claude Opus 5 review and the head-to-head in Sonnet 5 vs Opus 5 for real workloads. For the complete cost picture across the whole lineup, including Haiku 4.5, read Claude 5 family pricing explained.