Claude Sonnet 5 — What Is New and Cheaper Until August 31

Anthropic released Claude Sonnet 5 today, June 30, 2026, and the headline is simple: near-Opus coding quality at Sonnet prices — and, until August 31, at below Sonnet prices. The new model launches at an introductory $2 per million input tokens and $10 per million output tokens, then moves to the standard $3/$15 on September 1. If you have been waiting for a reason to re-benchmark your stack, this is it, and the clock is already running.

TL;DR:

• Claude Sonnet 5 (claude-sonnet-5) launched June 30, 2026 with near-Opus coding quality, a 1M-token context window, and 128K max output.

• Intro pricing is $2/$10 per MTok through August 31, 2026, then $3/$15 — the same list price as Sonnet 4.6.

• A new tokenizer produces roughly 30% more tokens for the same text, so “same price as 4.6” is not the same bill. The intro window offsets that; September 1 is when the real math starts.

• Adaptive thinking is on by default, the full effort range (low → xhigh) is supported, and temperature/top_p/top_k and budget_tokens are gone — plan a small migration, not a drop-in swap.

• It is the new default model on claude.ai Free and Pro, and available on Max, Team, and Enterprise.

Update, July 12, 2026: this article was published on launch day and has been updated since. Two weeks in, the rollout has held steady: Sonnet 5 is serving as the default on Free and Pro, the intro pricing window is confirmed through August 31, and the migration notes below reflect what we have seen moving real workloads over.

Claude Sonnet 5 at a glance

Here is the spec sheet next to the model it replaces. Sonnet 4.6 remains available on the API, so nothing breaks today — but the two models are further apart than a minor version bump suggests.

Spec	Claude Sonnet 5	Claude Sonnet 4.6
API model ID	claude-sonnet-5	claude-sonnet-4-6
Release date	June 30, 2026	February 2026
API price (in/out per MTok)	$2 / $10 until Aug 31, 2026, then $3 / $15	$3 / $15
Context window	1M tokens	1M tokens
Max output	128K tokens	128K tokens
Thinking	Adaptive, on by default	Extended thinking with budgets
Effort parameter	low / medium / high / xhigh	Partial range
Tokenizer	New (~30% more tokens per text)	Previous generation
Sampling params (temperature etc.)	Rejected	Accepted

The launch landed on a busy day: June 30 was also the day US export controls on Claude Fable 5 were lifted, with Fable restored globally on July 1. Anthropic clearly wanted its mid-tier workhorse refreshed before the flagship came back online. For where each model sits in the lineup, our Claude models comparison covers the full family.

What is new in Claude Sonnet 5 vs Sonnet 4.6

Near-Opus coding quality

Anthropic’s positioning in the launch announcement is that Sonnet 5 approaches Opus-tier quality on coding tasks, and in my first two weeks of daily use that claim mostly holds. On multi-file refactors and test-driven agent loops in Claude Code, Sonnet 5 completes tasks I would previously have escalated to Claude Opus 4.8. It is not Opus — on long-horizon agentic work that runs for hours, Opus still recovers from dead ends more gracefully — but for the 80% of coding work that is bounded and well-specified, the gap has become hard to notice.

The full effort range, including xhigh

Sonnet 5 supports the complete effort parameter range: low, medium, high, and xhigh. Previous Sonnet models did not get the top of that range. In practice, xhigh on Sonnet 5 is the setting that produces the near-Opus results — it spends noticeably more tokens reasoning before answering. Low effort, meanwhile, makes Sonnet 5 a credible replacement for some Haiku workloads where you want better judgment per request.

Adaptive thinking by default

Thinking is no longer something you configure with a token budget. Sonnet 5 ships with adaptive thinking on by default: the model decides per request how much reasoning a task needs. The old budget_tokens knob is gone entirely (more on that in the migration section). For most applications this is strictly better — easy prompts stay fast, hard prompts get the reasoning they need without you tuning anything.

A new tokenizer — and yes, it changes your bill

Sonnet 5 uses an updated tokenizer, similar to the change Anthropic introduced with Opus 4.7. The same input maps to more tokens — Anthropic’s own figure is roughly 1.0–1.35× depending on content type, and on typical mixed code-and-prose workloads I have measured around 30% more tokens than Sonnet 4.6 for identical text. This deserves its own section, because it is the detail most coverage glosses over.

Speed, in practice

Anthropic markets Sonnet as the best speed-to-intelligence balance in the lineup, and Sonnet 5 keeps that character. At default settings it streams responses noticeably faster than Opus 4.8 on comparable prompts — I would call it comfortably interactive where Opus sometimes makes you wait. Crank effort to xhigh and that advantage narrows, because the model spends real time reasoning; the point is that you now choose per request where on that curve to sit, instead of choosing a different model. For agent workloads that mix trivial and hard steps, that single knob replaces a lot of routing logic.

The tokenizer math: what Sonnet 5 actually costs

Per-token prices only tell you half the story when the tokenizer changes underneath them. Run the numbers on a workload that consumed 100M input tokens per month on Sonnet 4.6:

• Sonnet 4.6: 100M tokens × $3/MTok = $300/month
• Sonnet 5, intro pricing (through Aug 31): ~130M tokens × $2/MTok = $260/month — you come out ahead
• Sonnet 5, standard pricing (from Sep 1): ~130M tokens × $3/MTok = $390/month — roughly 30% more than 4.6, despite the identical list price

Anthropic has been reasonably upfront that the intro pricing is set so the transition is roughly cost-neutral, and the official what’s-new page documents the tokenizer multiplier. But the honest framing is this: after August 31, Sonnet 5 is an effective price increase over Sonnet 4.6 for the same text, paid for (in theory) by better output quality and fewer retries. Whether that trade works depends on your workload — which is exactly why the intro window matters. Prompt caching softens the blow considerably, since cache reads bill at roughly a tenth of the input price; our Claude API pricing guide walks through the caching multipliers.

Intro pricing ends August 31, 2026 — use the window

The $2/$10 introductory rate runs through August 31, 2026. That gives you a two-month evaluation period where switching from Sonnet 4.6 is cheaper even after the tokenizer penalty. Treat it as exactly that:

1. Move a representative slice of production traffic to claude-sonnet-5 now, while mistakes are cheap.
2. Log actual token counts per request, not estimated ones — your tokenizer multiplier is workload-specific (code skews higher than prose in my measurements).
3. Before September 1, recompute your monthly cost at $3/$15 with the observed multiplier and decide deliberately whether each workload stays on Sonnet 5, drops back to 4.6, or moves down to Haiku.

Teams that skip step 3 will get a quiet 30% bill increase on September 1 and wonder why. Put a calendar reminder on August 25. It is the least glamorous advice in this article and the most valuable.

Migration notes for API users

Sonnet 5 is not a drop-in swap if your integration dates back a couple of model generations. Two things will break:

Sampling parameters are rejected

Requests that include temperature, top_p, or top_k return a validation error on Sonnet 5, as they already do on Opus 4.7 and later. If your request builder sets temperature: 0 “for determinism” — a very common pattern in older codebases — strip it before switching the model ID. Grep for all three parameters; I found two forgotten top_k settings in a service I was sure was clean.

budget_tokens is gone

The fixed thinking budget is removed on current models. If you were sending thinking: {"type": "enabled", "budget_tokens": 8000}, replace it with adaptive thinking — or send nothing, since adaptive is the default:

{
  "model": "claude-sonnet-5",
  "max_tokens": 4096,
  "thinking": {"type": "adaptive"},
  "effort": "high",
  "messages": [{"role": "user", "content": "..."}]
}

Where you used a large thinking budget to force deeper reasoning, effort: "xhigh" is the new lever. Where you used a small budget to cap latency, use effort: "low". The mapping is not exact, but it covers every use of budget_tokens I have seen in the wild.

Who should switch to Claude Sonnet 5

• From Sonnet 4.6: yes, and soon. The quality jump is real, the migration is small, and switching during the intro window means you evaluate at a discount. Just do the September math before September does it to you.
• From Opus 4.8: selectively. At intro pricing, Sonnet 5 costs 40% of Opus per input token ($2 vs $5) and its coding quality is close enough that bounded tasks — code review, single-feature implementation, structured extraction — often cannot justify the Opus premium anymore. Keep Opus for long-running agents and genuinely hard reasoning. The trade-offs mirror the ones in our Opus vs Sonnet deep comparison, updated for the new generation.
• From Haiku 4.5: mostly no. Haiku remains 2–3× cheaper per token even at intro pricing, and the tokenizer gap widens that. Switch only the requests where Haiku’s quality ceiling is actually costing you retries.

Availability on claude.ai plans

Claude Sonnet 5 rolled out across every consumer tier on launch day. It is now the default model on the Free plan and on Pro ($20/month, or $17/month billed annually), and it is available in the model picker on Max, Team, and Enterprise. It also works in Claude Code within your plan’s usage budget, where the effort control pairs nicely with agentic sessions — xhigh for the gnarly refactor, low for the boilerplate.

For Free users this is the largest quiet upgrade in a while: the default model went from “good mid-tier” to “near-flagship coding quality” overnight, with no plan change required. If you are choosing between tiers, the full guide section covers plans, features, and limits in detail.

One practical note for Pro subscribers: because Sonnet 5 is now the default, your existing chats and Projects simply start using it — there is nothing to configure. If you had manually pinned Sonnet 4.6 in the model picker at some point, that selection persists, so check the picker if your responses have not changed character since June 30. In Claude Code, sessions on the default model picked up Sonnet 5 on launch day as well, and the usage budget accounting works the same as before.

What people get wrong about Sonnet 5

• “Same price as 4.6, so the upgrade is free.” Nominal price, yes. Effective cost after August 31 is ~30% higher for the same text because of the tokenizer. Budget for it.
• “My requests started failing — Sonnet 5 is broken.” Check for temperature, top_p, top_k, or budget_tokens in the payload. The model rejects them by design; the error message tells you exactly which parameter to remove.
• “Intro pricing will probably get extended.” Maybe, but nothing Anthropic has published says so. Plan against August 31, 2026.
• “Sonnet 5 replaces Opus.” It narrows the gap on coding; it does not close it on long-horizon agentic work, and Opus 4.8 remains the recommendation for the hardest tasks.
• “Sonnet 4.6 is dead.” It is still an active API model at $3/$15. If your workload is tokenizer-sensitive and quality-insensitive, staying put for now is a legitimate choice.

FAQ

Bottom line

Claude Sonnet 5 is the rare release where the launch-week advice is genuinely time-sensitive: the model is better, the migration is small, and the intro pricing makes the next two months the cheapest evaluation window you will get. Move a workload over, measure your real tokenizer multiplier, and decide before September 1. This article tracks the release itself; for the long-term, hands-on verdict as it develops, see our evergreen Claude Sonnet review.