Claude Sonnet 5: the new default, at a discount until August 31
In brief
Claude Sonnet 5 launched June 30, 2026 and is now the default model on Free and Pro. It runs at introductory pricing of $2/$10 per million tokens through August 31, then goes to $3/$15. Three things break when you migrate from Sonnet 4.6, and a new tokenizer means the same text costs about 30% more tokens.
Contents
Claude Sonnet 5 (claude-sonnet-5) launched on June 30, 2026. It is now the default model on Free and Pro plans, and it's available on Max, Team, Enterprise, Claude Code, and the Claude API.
| Claude Sonnet 5 | |
|---|---|
| Model ID | claude-sonnet-5 |
| Context window | 1M tokens |
| Max output | 128k tokens |
| Thinking | Adaptive, on by default |
| Price | $2 / $10 per MTok |
| Priority Tier | Not available |
The pricing clock was cancelled
Update, August 10, 2026: the $2 / $10 rate is now permanent. It launched as introductory pricing due to expire on August 31, 2026, with a scheduled rise to $3 / $15 on September 1. Anthropic cancelled that increase. The standard price is $2 input / $10 output per million tokens, with no end date.
This is a real change to agent economics, not a footnote. Sonnet 4.6 ran at $3 / $15, so Sonnet 5 is a permanent 33% cut against the model it replaces — and Sonnet 5 is what most production agent fleets run on. If you built a cost model that assumed a 50% increase in September, you have just recovered a third of your headroom.
It also means the Haiku 4.5 versus Sonnet 5 decision is not what it used to be. Sonnet is now 2× Haiku rather than 3.75×, which weakens the standard argument for dropping a tier on high-volume work.
Three things break when you migrate from Sonnet 4.6
1. Adaptive thinking is on by default
Same change as Opus 5: requests now run with thinking unless you say otherwise. Because max_tokens caps total output including thinking tokens, revisit that value on anything you migrate.
2. Manual extended thinking is gone
# 400 error on Sonnet 5
thinking={"type": "enabled", "budget_tokens": 10000}
Manual thinking budgets were deprecated on Sonnet 4.6 and are now removed. Passing thinking: {type: "enabled", budget_tokens: N} returns a 400 error. Adaptive thinking replaces it — the model chooses its own depth, and you steer with the effort parameter instead of a token budget.
3. Sampling parameters return an error
# 400 error on Sonnet 5
temperature=0.7
top_p=0.9
top_k=40
Setting temperature, top_p, or top_k to any non-default value returns a 400. This catches people out, because temperature-tuning is one of the oldest habits in LLM work and a lot of older code sets temperature=0 reflexively for "deterministic" output.
If your codebase has a shared client wrapper that sets temperature on every call, that wrapper will break every Sonnet 5 request. Check it before you migrate, not after.
The tokenizer change is a real cost line
Sonnet 5 uses the tokenizer introduced with Opus 4.7. Compared to models before Opus 4.7, the same text produces roughly 30% more tokens.
This is not a price change, but it lands on your bill like one. A prompt that counted 10,000 tokens on Sonnet 4.6 may count around 13,000 on Sonnet 5. The exact increase depends on your content — code, non-English text, and structured data all shift differently.
Two things follow:
- Re-measure, don't estimate. Use the token counting API with
model: "claude-sonnet-5"against your real prompts. A blanket 30% assumption will be wrong for your specific workload in one direction or the other. - Check your context headroom. If you were running near a context limit on Sonnet 4.6, the same input may not fit the same way. The 1M window gives most workloads plenty of room, but anything that was already tight deserves a look.
Combined with the introductory pricing, the net effect through August is still cheaper than Sonnet 4.6 for most workloads. From September, run the numbers again with the new tokenizer counts and the $3/$15 rate.
No Priority Tier
Sonnet 5 supports the same tools and platform features as Sonnet 4.6 with one exception: Priority Tier is not available. If you pay for Priority Tier to get guaranteed capacity during traffic spikes, that does not extend to Sonnet 5. This is the one reason a production workload might reasonably stay on Sonnet 4.6 for now.
What actually got better
Anthropic reports substantial gains over Sonnet 4.6 in multi-step reasoning, tool use, coding, and knowledge work. The specific thing early testers highlighted: completing complex agentic workflows end to end, and self-verifying without being told to.
That last behavior carries the same caveat as Opus 5 — if your prompts contain instructions telling the model to verify its work, they were written for an older model and now cause redundant checking. Remove them.
Migration checklist
- Search your code for
temperature,top_p, andtop_k. Remove non-default values, especially in shared client wrappers. - Search for
budget_tokens. Remove it. - Re-run token counting against real prompts with
model: "claude-sonnet-5". - Recheck
max_tokensnow that thinking counts against it. - Confirm nothing you're migrating depends on Priority Tier.
- Put September 1 in the calendar as the date your Sonnet 5 spend rises 50%.
Related reading
- Claude Opus 5 — the new flagship
- Choosing the right Claude model — Sonnet vs Opus vs Haiku
- Claude cost optimization — controlling spend
- Minimising token usage — matters more under the new tokenizer
Source: Claude Platform release notes, June 30, 2026, and Introducing Claude Sonnet 5.