Claude Sonnet 5.5 is out. Should it replace Opus 5.5 for your work?
Anthropic released Sonnet 5.5 on September 28, 2026. Here are the official terms, the vendor’s own scores, where you can use it today, and which pre-release leaks turned out to be right.
Same price as Sonnet 5, half the token price of Opus 5.5. Try it first on routine, well-scoped work; keep Opus 5.5 for open-ended problems that need judgment.
That split is Anthropic’s own positioning, not a test we ran. Check it on one real task of yours before switching a whole workflow.
Read Anthropic’s announcement (opens in a new tab)What is confirmed
- Released
- September 28, 2026Official Source (opens in a new tab)
- API model ID
- claude-sonnet-5-5 (Bedrock: anthropic.claude-sonnet-5-5)Official Source (opens in a new tab)
- API price
- $2 input / $10 output per 1M tokens; cache reads $0.20; 5-minute cache writes $2.50. Same as Sonnet 5.Official Source (opens in a new tab)
- Context and output
- 1M-token context; 128K max output (300K on the Batch API with a beta header)Official Source (opens in a new tab)
- Default effort
- Medium in the Claude apps and Claude Code; High on the Claude Platform APIOfficial Source (opens in a new tab)
- Speed and task cost
- 30%+ faster output than Sonnet 5; up to 30% lower cost per task in Anthropic’s testingVendor claim Source (opens in a new tab)
- Haiku 5.5
- “In the coming weeks.” No date, price or model card yet.Official, undated Source (opens in a new tab)
Close to Opus 5.5 on paper. Far ahead of Sonnet 5.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0Agentic coding | 70.6% | 10.3% | 66.4% |
| FrontierCode 1.1 (main)Agentic coding | 52.1% Xhigh · 46.2% Max | 42.4% | 54.4% |
| CursorBench 4.0Agentic coding | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1Knowledge work (Elo) | 1844 | 1449 | 1846 |
| AA-Briefcase v1.1Long knowledge work (Elo) | 1811 | 1359 | 1822 |
| Humanity’s Last ExamReasoning, with tools | 64.5% | 54.9% | 67.7% |
| OSWorld 2.1Computer use | 80.1% | 57.0% | 81.8% |
| ChartographyChart reading, no tools | 61.6% | 15.6% | 64.4% |
These are Anthropic’s numbers, not independent tests. Opus 5.5’s Terminal-Bench score is at Xhigh effort. On FrontierCode, Sonnet 5.5 scored lower at Max than at Xhigh; Anthropic says Max more often ran a multi-agent review that timed out or edited beyond the task. Anthropic also says that in its own and testers’ use, Opus 5.5 “remains clearly stronger” at complex, open-ended work.
56 on the Artificial Analysis Intelligence Index v4.3.2 (Adaptive reasoning, max effort, default fallback), 3rd of 216 models. Opus 5.5 scores 58 on the same index. This is Artificial Analysis’s measurement, not ours.
What people are saying—and making
Better at revision than he expected—and his team splits on whether a mid-tier model still has a place.
Shipper says its writing improved dramatically in his testing for Every, beating GPT-6 Astra at revision tasks though Astra stays his favorite writer. A colleague prefers it for quick iterative coding and design; two others say they no longer need a mid-tier model. Personal reports from an early-access publisher, not a controlled study.
Editorial paraphrase · Sonnet 5.5 vs Opus 5.5 and GPT-6 Astra · context checked 2026-09-29. Original-source availability may vary.
- Watch the 30-second video at its source
- Note what a single chat session is claimed to produce
- Ask for the code and one changed requirement before reusing it
A candy-colored 3D run, made in one early Sonnet 5.5 chat
A 30-second multi-camera animation: a small blob character wakes up in a grey world, the scene floods with color, and a countdown sends it racing. The creator says one Sonnet 5.5 chat session made it.
Published 2026-09-28 · Free case note
One day after launch, few public projects say which model made them. We add work only when the creator states Sonnet 5.5 was used, and we keep early-access builds labelled.
Available, rolling out, or not yet
Claude apps (web, iOS, Android)
Anthropic says anyone can chat with Sonnet 5.5 on Claude.ai. Message limits depend on your plan and are not listed on the launch page.
Source (opens in a new tab)Claude Code
Version 2.1.284 (September 28) added it as the default Sonnet on the Anthropic API. Run claude --version and update if you are older.
Source (opens in a new tab)Claude API, Bedrock, Google Cloud, Microsoft Foundry
Same model on all four; Claude Platform on AWS too.
Source (opens in a new tab)GitHub Copilot
Pro, Pro+, Max, Business and Enterprise. Gradual rollout; billed at provider list price under usage-based billing.
Source (opens in a new tab)Cursor
In Cursor’s model list at $2 / $10 per 1M tokens, cache write $2.50, cache read $0.20.
Source (opens in a new tab)Kiro
Kiro’s model table lists Sonnet 5 and, since September 28, Opus 5.5 (paid tiers, 2.0x credits). No Sonnet 5.5 when we checked.
Source (opens in a new tab)A rollout can reach accounts at different times. If it is missing from your picker, update the app or check again later; availability in one product does not mean it is in another.
Four changes that can surprise you
Thinking is on by default
A request with no thinking field now runs adaptive thinking. To keep up-front thinking off, send thinking: {"type": "between_tools"}; "disabled" is no longer accepted.
Forced tool choice returns an error
So do thinking budgets, non-default temperature/top_p/top_k and assistant prefill. Check these five before you swap the model ID.
Text between tool calls moves into thinking blocks
An app that streams that text to users goes quiet between tool calls until it sets a display value or uses between_tools.
High-risk cyber requests fall back to Sonnet 5
It is the first Sonnet with Opus-style cyber safeguards. Routine bug fixing is unaffected; flagged requests visibly switch model.
What the pre-release leaks got right
Before launch, several reports described partner builds of Sonnet 5.5. We compare each claim with the official documentation after release. A correct leak is still not confirmation until the vendor publishes.
Release “as soon as early next week, possibly before Monday morning ends”
“Lyra” on X, via TestingCatalog (Sep 27)
Released Monday, September 28.
Leak report (opens in a new tab)Forced tool use retired
“Lyra” on X, via TestingCatalog (Sep 27)
Forced tool choice returns a 400 error on Sonnet 5.5.
Leak report (opens in a new tab)1M context, 128K output, $2 / $10
“Lyra” on X, via TestingCatalog (Sep 27)
These were already Sonnet 5’s published terms; matching them told readers little.
Leak report (opens in a new tab)Launch around September 30 or October 1
An X account tracking release patterns, via Startup Fortune (Sep 27)
It shipped two days earlier.
Leak report (opens in a new tab)Questions readers ask
When was Claude Sonnet 5.5 released?
September 28, 2026, on the Claude apps, Claude Code, the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Official source (opens in a new tab)How much does Sonnet 5.5 cost?
On the API, $2 per million input tokens and $10 per million output tokens, with $0.20 cache reads—the same as Sonnet 5. In the Claude apps it is part of your plan’s usage, not billed per token.
Official source (opens in a new tab)Is Sonnet 5.5 better than Opus 5.5?
No, by Anthropic’s own account. Its scores come close to Opus 5.5 on several benchmarks, but Anthropic says Opus 5.5 remains clearly stronger at complex, open-ended work. Sonnet 5.5 has half the input and output token price.
Official source (opens in a new tab)What is the Sonnet 5.5 context window?
1M tokens, with up to 128K output tokens per response.
Official source (opens in a new tab)Can I use Sonnet 5.5 in Kiro?
Not as of September 29, 2026: Kiro’s model table lists Sonnet 5 and Opus 5.5, but not Sonnet 5.5.
Official source (opens in a new tab)When is Haiku 5.5 coming?
Anthropic says “in the coming weeks.” There is no official date.
Official source (opens in a new tab)Facts were checked on September 29, 2026 against the linked pages. Benchmark and speed figures are Anthropic’s and its customers’ claims. We have not yet run a controlled Sonnet 5.5 comparison of our own.