On September 28, 2026, Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family after Claude Opus 5.5 launched six days earlier. Anthropic describes it as a "clear upgrade" over Claude Sonnet 5: it runs over 30% faster, costs up to 30% less per task, and posts sharply higher scores on agentic coding tests, while keeping Sonnet 5's list price unchanged.
What's new in Claude Sonnet 5.5
Anthropic positions Sonnet 5.5 as the fast, lower-cost complement to Opus 5.5: strongest on well-scoped everyday work, bug fixes, and polished documents, slides, and spreadsheets, while Opus 5.5 remains its pick for complex work that needs sustained judgment (Anthropic). Anthropic also says a Claude Haiku 5.5 model is coming "in the coming weeks" to complete the 5.5 lineup.
The model is a large step up from Sonnet 5 on coding and long-horizon tasks. Anthropic reports Sonnet 5.5 as the first Sonnet model to beat Pokemon Red working only from screenshots, and says early testers described it as writing more clearly and needing fewer back-and-forth steps than Sonnet 5, in part because it batches tool calls together more efficiently (Anthropic).
Anthropic
AWS
Google Cloud
MicrosoftHow it compares on benchmarks
Anthropic published launch-day scores against Sonnet 5 and its own Opus 5.5. Higher is better in every row.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 70.6% | 10.3% | 66.4% |
| CursorBench 4.0 (agentic coding) | 55.5% | 34.1% | 57.8% |
| FrontierCode 1.1, Main set (Max effort) | 46.2% | 42.4% | 54.4% |
| GDPval-AA v2.1 (real-world knowledge work) | 1844 | 1449 | 1846 |
| AA-Briefcase v1.1 (long-horizon knowledge work) | 1811 | 1359 | 1822 |
| Humanity's Last Exam, with tools | 64.5% | 54.9% | 67.7% |
| OSWorld 2.1 (computer use) | 80.1% | 57.0% | 81.8% |
| Chartography (chart recognition, no tools) | 61.6% | 15.6% | 64.4% |
Source: Anthropic, Claude Sonnet 5.5 launch page. Figures are self-reported by the vendor and use Anthropic's own evaluation harnesses; they have not been independently audited.
The biggest jump is Terminal-Bench 4.0, where Sonnet 5.5 goes from 10.3% to 70.6%, ahead of Opus 5.5's 66.4%. On GDPval-AA and AA-Briefcase, both tests of real-world professional work, Sonnet 5.5 lands within a few points of Opus 5.5 and roughly 350 to 450 points ahead of Sonnet 5. Oddly, Anthropic reports Sonnet 5.5 scoring higher on FrontierCode 1.1 at Xhigh effort (52.1%) than at its top Max effort (46.2%), which it attributes to a code-review skill that, at Max, led to timeouts or edits beyond the task's scope (Anthropic). Anthropic is explicit that these gains narrow, rather than close, the gap to Opus 5.5, which it says stays clearly ahead on complex, open-ended work that needs sustained judgment.
Sonnet 5.5 doesn't advance the frontier of our models' capabilities, so our alignment assessment focused on a targeted set of risks that apply to models of any capability level.
Anthropic, Claude Sonnet 5.5 launch page
Pricing and effort levels
Sonnet 5.5 carries the exact same list price as Sonnet 5: $2 per million input tokens, $10 per million output tokens, $0.20 per million tokens for cache reads, and $2.50 per million tokens for standard five-minute cache writes (Anthropic pricing docs; Anthropic launch page). That $2/$10 rate was originally an introductory price for Sonnet 5 that Anthropic has since made permanent, and Sonnet 5.5 inherits it unchanged.
| Price per million tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache write (5 min) | $2.50 | $5 |
| Cache read | $0.20 | $0.20 |
Source: Anthropic pricing documentation.
Because the price per token has not moved, the "up to 30% less per task" saving Anthropic claims comes entirely from Sonnet 5.5 needing fewer tokens and fewer steps to reach the same result. Anthropic's customer quotes echo this: Balyasny Asset Management said Sonnet 5.5 used roughly 121,000 tokens per answer on its finance test suite versus 497,000 for Sonnet 5, and Slack reported about 14% fewer output tokens on its Slackbot evaluations (Anthropic launch page).
Sonnet 5.5 has a 1 million token context window and a 128,000 token maximum output on the standard Messages API (Anthropic model overview). It supports five effort levels, low, medium, high, xhigh, and max, which trade cost and latency against reasoning depth. The default is high effort on the Claude Platform API and medium effort in the Claude apps and Claude Code.
What changes for developers
Anthropic's migration guide lists several breaking changes for code built on Sonnet 5. The thinking: {"type": "disabled"} setting used to turn off up-front reasoning now returns an error; it is replaced by a new between_tools setting that keeps thinking off up front while still returning progress updates between tool calls. Forcing a specific tool via tool_choice is also rejected outright, in favor of auto combined with strict tool schemas. Thinking blocks are now bound to both the account that produced them and the conversation prefix before them, so editing earlier turns and replaying an older thinking block can trigger an error on newer accounts.
Safety changes
Anthropic says Sonnet 5.5 matches or improves on Sonnet 5 across most measures in its automated behavioral audit, a roughly 1,850-scenario test of alignment and misuse resistance. Because its cybersecurity capabilities have grown enough to approach Opus 5's, Sonnet 5.5 is the first Sonnet model to ship with the same real-time cyber safeguards Anthropic built for its most capable models; higher-risk cybersecurity requests visibly fall back to Sonnet 5 instead. Its biology safeguards are unchanged from Sonnet 5's. Sonnet 5.5 is also the first Sonnet model to launch with safety classifiers aimed at reasoning-extraction attacks, a defense against distillation attempts that try to copy a model's internal reasoning at scale (Anthropic launch page).
Getting started
Claude Sonnet 5.5 is available now on the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure, using the model ID claude-sonnet-5-5, with zero data retention offered as on Opus 5.5 and Sonnet 5 (Anthropic launch page). On Metir, Sonnet 5.5 has replaced Sonnet 5 as the Claude Sonnet option in chat on both web and mobile, so picking it from the model menu needs no separate setup.
Sources:
- Anthropic: Introducing Claude Sonnet 5.5
- Anthropic: Claude Sonnet 5.5 model overview
- Anthropic: Migrating to Claude Sonnet 5.5
- Anthropic: Claude pricing
- Metir AI blog: Claude Opus 5.5 launch coverage
Image credits
- Hero: Dario Amodei, co-founder and CEO of Anthropic, speaking at TechCrunch Disrupt in 2023. Source: Wikimedia Commons, photo by TechCrunch, licensed CC BY 2.0.
- In-body: Dario Amodei (third from left) meeting Japanese Prime Minister Sanae Takaichi in 2025. Source: Wikimedia Commons, photo by the Cabinet Secretariat, Government of Japan, licensed CC BY 4.0.