Claude Sonnet 5
Claude Sonnet 5 is Anthropic's best mid-price model and the default in the free and Pro Claude apps. It launched on 30 June 2026 at $2 per million input tokens and $10 per million output tokens. That is a third cheaper per token than Sonnet 4.6 ($3/$15). Anthropic says it performs close to Opus 4.8 on reasoning, tool use, coding and knowledge work.
It has a 1M-token context window, 128K max output and a January 2026 knowledge cutoff. There is one catch on price: Sonnet 5 uses a new tokenizer that turns the same text into roughly 1.0 to 1.35 times as many tokens. So the real saving over Sonnet 4.6 is smaller than the sticker price suggests. Anthropic has said a Sonnet 5.5 is coming within weeks of the 22 September Opus 5.5 launch.
Claude Sonnet 5 is the sensible default for most everyday and high-volume work. Anthropic's figures show it catching or passing Opus 4.8 on several tests. It scores 80.4% on Terminal-Bench 2.1 and 1618 on GDPval-AA v2 (Opus 4.8: 1615). It stays close on Humanity's Last Exam with tools (57.4% vs 57.9%). It does all this at $2/$10 per million tokens, and it is fast.
It is a clear step up from Sonnet 4.6: +5.1 points on SWE-bench Pro, +13.4 on Terminal-Bench 2.1 and +223 Elo on GDPval-AA v2. Free and Pro users of the Claude app get it by default.
Who should pick it: developers who need a cheap, fast model for coding help, agents, support bots and document work, and anyone on a free or Pro plan.
Who should not: teams doing the hardest multi-hour coding or research. Opus 5.5 scores far higher on independent tests (Artificial Analysis 54 at high effort vs Sonnet 5's 32) for only twice the price. If you still use temperature settings, fix that code first; Sonnet 5 rejects non-default values.
With Sonnet 5.5 announced as coming soon, avoid long commitments tied to this exact version.
Score breakdown
Best for
- Everyday chat and writing in the Claude app
- High-volume API work such as support bots and data extraction
- Coding assistants and agents on a budget
- Sub-agents working under a bigger Opus model
What we like
- Close to Opus 4.8 on Anthropic's figures at $2/$10 per 1M tokens
- Beats Opus 4.8 on GDPval-AA v2 knowledge work (1618 vs 1615)
- Fast, and the default model on Claude Free and Pro
- 1M-token context and 128K output
- Anthropic reports lower rates of bad behaviour and better prompt-injection resistance than Sonnet 4.6
Watch out for
- New tokenizer uses up to 1.35x more tokens, eating into the headline price cut
- Far behind Opus 5.5 on independent tests (Artificial Analysis 32 vs 54 at high effort)
- Rejects non-default temperature, top_p and top_k values
- Sonnet 5.5 is already announced, so this version may not stay current for long
Specs at a glance
| API model ID | claude-sonnet-5 |
|---|---|
| Amazon Bedrock ID | anthropic.claude-sonnet-5 |
| Other platforms | Google Cloud, Microsoft Foundry, Claude Platform on AWS |
| Context window | 1,000,000 tokens |
| Max output | 128K tokens (300K on Batch API with beta header) |
| Input / output | Text and images in, text out |
| Thinking | Adaptive, on by default (can be disabled) |
| Default effort | high |
| Comparative latency | Fast |
| Reliable knowledge cutoff | January 2026 |
| Tokenizer | New; about 1.0 to 1.35x the tokens of Sonnet 4.6 for the same text |
| Sampling parameters | Non-default temperature, top_p or top_k return an error |
| Retirement | Not sooner than 30 Jun 2027 |
| Predecessor | Claude Sonnet 4.6 |
Benchmarks
Benchmarks are standard tests. Vendor-run results are marked as such; independent results are preferred where they exist.
| Benchmark | Score | Source | Note |
|---|---|---|---|
| SWE-bench Pro (agentic coding) | 63.2% | Anthropic via DataCamp/Vellum | Sonnet 4.6: 58.1%; Opus 4.8: 69.2% |
| Terminal-Bench 2.1 | 80.4% | Anthropic via DataCamp/Vellum | Sonnet 4.6: 67.0% |
| Humanity's Last Exam (no tools / with tools) | 43.2% / 57.4% | Anthropic via DataCamp | Opus 4.8: 49.8% / 57.9% |
| OSWorld-Verified (computer use) | 81.2% | Anthropic via DataCamp/Vellum | Sonnet 4.6: 78.5%; Opus 4.8: 83.4% |
| GDPval-AA v2 (knowledge work, Elo) | 1618 | Anthropic via DataCamp/Vellum | Sonnet 4.6: 1395; Opus 4.8: 1615 |
| CursorBench | 57% | Cursor via Vellum | Sonnet 4.6: 49% |
| Artificial Analysis Intelligence Index | 32 (high) / 38 (max) | Artificial Analysis |
Pricing
| Plan / tier | Price | Notes |
|---|---|---|
| API input | $2 per 1M tokens | |
| API output | $10 per 1M tokens | |
| Cache write (5 min / 1 h) | $2.50 / $4 per 1M tokens | |
| Cache read | $0.20 per 1M tokens | |
| Batch API | 50% off ($1 / $5) | |
| Claude apps | Default model on Free and Pro | Also on Max, Team and Enterprise |
Sonnet 5 vs Sonnet 4.6 vs Opus 4.8
Anthropic's launch figures, as reported by DataCamp and Vellum:
| Test | Sonnet 5 | Sonnet 4.6 | Opus 4.8 |
|---|---|---|---|
| SWE-bench Pro | 63.2% | 58.1% | 69.2% |
| Terminal-Bench 2.1 | 80.4% | 67.0% | not comparable* |
| Humanity's Last Exam (tools) | 57.4% | 46.8% | 57.9% |
| OSWorld-Verified | 81.2% | 78.5% | 83.4% |
| GDPval-AA v2 | 1618 | 1395 | 1615 |
| Price (in / out) | $2 / $10 | $3 / $15 | $5 / $25 |
*The two reports give different Opus 4.8 figures for Terminal-Bench 2.1 (74.6% and 82.7%), so we leave it out.
The tokenizer catch
A tokenizer splits text into the chunks a model counts and bills. Sonnet 5 uses a new one. Anthropic says the same input now becomes about 1.0 to 1.35 times as many tokens, depending on the content.
What that means for your bill, using list prices:
- Sonnet 4.6 input: $3 per million tokens.
- Sonnet 5 input: $2 per million tokens, but up to 1.35x more tokens, so up to about $2.70 for the same text.
Sonnet 5 is still cheaper, but by about 10% to 33%, not a flat 33%. Measure real token counts with the API's token counter before you estimate savings.
Developer changes from Sonnet 4.6
Anthropic calls Sonnet 5 a drop-in upgrade, with three behaviour changes:
- Adaptive thinking is on by default. The model decides how much to reason. You can still disable it.
- Manual thinking budgets are gone.
budget_tokensnow returns an error. Use theeffortsetting instead. - No custom sampling. Setting
temperature,top_portop_kto non-default values returns an error.
It also brings higher-resolution image input and cyber safeguards that are on by default.
- Count tokens before you budget. Run a sample of your real prompts through the token-counting endpoint, because the new tokenizer can add up to 35%.
- Use Sonnet 5 as a worker under Opus 5.5. Let Opus plan and check, and let Sonnet handle the many smaller steps at half the price.
- Drop
efforttolowormediumfor chat and simple extraction. The default ishigh, which spends more tokens than those jobs need. - Delete any
temperatureortop_psettings and anybudget_tokensvalues from old Sonnet 4.x code before switching, or requests will fail. - On the free Claude plan, start a new chat for each new topic. Long threads resend the whole history and use up your limit faster.
Jargon explained
- Tokenizer
- The part of a model that splits text into tokens. A new tokenizer can change how many tokens, and so how much money, the same text costs.
- SWE-bench Pro
- A tough coding test where the AI must fix real bugs in real software projects, checked by the projects' own tests.
- GDPval-AA
- A test of real office work, such as reports and spreadsheets, scored like a chess rating (Elo), where higher is better.
- Temperature
- An older setting that made AI replies more random or more predictable. Sonnet 5 no longer lets you change it.
- Sub-agent
- A helper AI that a main AI hands smaller jobs to, often a cheaper and faster model.
Alternatives to consider
Claude Opus 5.5
Much stronger on hard reasoning and long agent tasks for twice the per-token price.
Claude Haiku 4.5
Half the price and the fastest Claude, for simple, high-volume tasks.
Claude Sonnet 4.6
The predecessor. Only worth keeping if your code depends on custom temperature settings.
Gemini 3.5 Flash
Google's fast, low-cost rival in the same price class.
GPT-6 Sol
OpenAI's rival model, launched on the same day as Opus 5.5.
Frequently asked questions
When was Claude Sonnet 5 released?
Anthropic released Claude Sonnet 5 on 30 June 2026.
How much does Claude Sonnet 5 cost?
$2 per million input tokens and $10 per million output tokens on the API. Cache reads cost $0.20 per million and the Batch API halves prices. It is the default model on the free and Pro Claude plans.
Is Sonnet 5 better than Opus 4.8?
Close, and ahead on some tests. On Anthropic's figures it edges Opus 4.8 on GDPval-AA v2 (1618 vs 1615) but trails on SWE-bench Pro (63.2% vs 69.2%) and OSWorld-Verified (81.2% vs 83.4%).
Is Sonnet 5 cheaper than Sonnet 4.6?
Yes, but by less than the price list suggests. Its new tokenizer uses about 1.0 to 1.35 times as many tokens for the same text, so real savings are roughly 10% to 33%.
Is there a Claude Sonnet 5.5?
Not yet. When it launched Opus 5.5 on 22 September 2026, Anthropic said Sonnet 5.5 would follow in the coming weeks.
Why does Sonnet 5 reject my temperature setting?
Sonnet 5 does not accept non-default temperature, top_p or top_k values. Remove them and steer style with your prompt instead.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- Introducing Claude Sonnet 5 (Anthropic)
- Claude Sonnet 5 model overview (Claude Platform Docs)
- Claude Sonnet 5: Features, Benchmarks, and Pricing (DataCamp)
- Claude Sonnet 5 Benchmarks Explained (Vellum)
- Models overview (Claude Platform Docs)
- Introducing Claude Opus 5.5 (Anthropic)
- LLM Leaderboard: Intelligence Index (Artificial Analysis)