Claude Sonnet 5 is the next generation of Anthropic's Sonnet family, released on June 30, 2026. The official announcement calls it "the most agentic Sonnet model yet," and it became the default model for Free and Pro plans at launch. This guide covers what changed, how pricing works, and what existing users and developers should watch for when upgrading — based on the official announcement and developer docs.
Key specs at a glance
| Item | Claude Sonnet 5 |
|---|---|
| Release date | June 30, 2026 |
| API model ID | claude-sonnet-5 |
| Context window (how much conversation and documents the model holds at once) | 1M tokens — both the default and the maximum; no smaller variant |
| Max output | 128k tokens |
| Pricing (per million tokens) | Intro $2 input / $10 output through Aug 31, 2026 → then $3 / $15 |
| Thinking | Adaptive thinking (the model decides on its own when to think deeply) on by default |
| Availability | All plans — default for Free and Pro; Claude Code and the API |
What actually improved
According to the official announcement, Sonnet 5 can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models. Compared with its predecessor Sonnet 4.6, it improves substantially on reasoning, tool use, coding, and knowledge work, and comes close to the higher-tier Opus 4.8 at a lower price. Anthropic also reports that safety assessments found an overall lower rate of undesirable behaviors than Sonnet 4.6.
Put simply: the "sensible mid-size sedan" of the lineup just inherited much of the top trim's driving performance. If you are deciding between models, see our model selection guide and the model comparison.
Pricing — and the tokenizer change that is easy to miss
API pricing runs at the introductory rate ($2 input / $10 output per million tokens) through August 31, then moves to the standard $3 / $15 — the same standard rate as Sonnet 4.6. The announcement also notes up to 90% savings with prompt caching (a discount for resending the same prompt prefix) and 50% with batch processing (grouping non-urgent requests).
One thing developers must know: per the official docs, Sonnet 5 uses a new tokenizer (the rule that splits text into tokens), and the same text produces approximately 30% more tokens. Even with unchanged per-token pricing, the cost of an equivalent request can differ; a max_tokens budget tuned for Sonnet 4.6 may truncate output; and the 1M window holds less text than before. The official guidance is to recount tokens against Sonnet 5 rather than reuse figures measured on earlier models.
Developer migration — three things that now return a 400
The docs describe Sonnet 5 as a drop-in upgrade for Sonnet 4.6 with exactly three behavior changes:
- Adaptive thinking is on by default. No configuration needed; the model thinks deeply when it decides it should.
- Manual extended thinking is removed. Requests with
thinking: {type: "enabled", budget_tokens: N}return a 400 error. - Sampling parameters are locked. Setting
temperature,top_p, ortop_kto non-default values returns a 400 error.
Also note that Priority Tier (a dedicated-capacity option) is not available on Sonnet 5. Aside from these, code that already runs on Sonnet 4.6 needs no other changes, per the docs. For choosing between API models, see the API model comparison guide.
How each plan gets it
- Free & Pro: Sonnet 5 is the default model — nothing to configure.
- claude.ai chat (paid plans): per the official help center, Sonnet 5 supports a 1M-token context window in chat on all paid plans; Opus 4.8 and peers support 500K in chat.
- Claude Code: requires v2.1.197 or later; the
sonnetalias resolves to Sonnet 5, and it always runs with the 1M window on the API. For costs, see our Claude Code pricing guide.
Cybersecurity safeguards — a first for Sonnet
Sonnet 5 is the first Sonnet-tier model with real-time cybersecurity safeguards enabled by default. Requests involving prohibited or high-risk cybersecurity topics may be refused — and on the API, a refusal comes back as a successful HTTP 200 response with stop_reason: "refusal", not an error. If you run automated pipelines, handle that response shape explicitly.
FAQ
How long does the introductory price last?
Through August 31, 2026. From September 1, the standard rate of $3 input / $15 output per million tokens applies.
Will my Sonnet 4.6 code work as is?
Mostly, yes. Remove manual extended thinking and non-default sampling parameters (both now return 400 errors), and revisit max_tokens and cost estimates given the roughly 30% higher token counts.
Is there an extra charge for the 1M context window?
No. Per the official docs, 1M is the default for Sonnet 5 — no beta header and standard pricing.
This article is based on Anthropic's official announcement and developer documentation (platform.claude.com, support.claude.com) as of July 4, 2026. Pricing, availability, and lineup details may change — please re-check the official docs before making important decisions.