Claude Sonnet 5.5

Early hands-on users praise how close it gets to Opus 5.5 at half the price, and its speed. They criticize token use at maximum effort.

Sonnet 5.5 is available even on Claude's free plan, and its API price matches Sonnet 5. You cannot download it.

Of six requests, one is a no: download. The rest are covered by a Claude subscription or the API.

You fix bugs and write code in Claude Code every day. On CursorBench, Sonnet 5.5 trails Opus 5.5 by two points, but costs half as much.

You build documents, spreadsheets and slides. On office tasks, GDPval-AA puts it nearly level with Opus 5.5: 1844 vs 1846.

You pay for the API per token. Rates are unchanged from Sonnet 5, and the model completes the work in fewer steps.

You need long autonomous work that requires fine judgment. Anthropic says plainly that Opus 5.5 is noticeably stronger there.

You want to run it at maximum effort. Artificial Analysis prices that task at $7.60, more than Opus 5.5.

Ways to run it are covered in chapter 4, pricing in chapter 6. Choosing between Sonnet and Opus is covered in chapter 8.

On well-defined tasks, Sonnet 5.5 works almost like Opus 5.5, only faster and cheaper. On the first evening, people tested this on games, simulations and code edits.

Bug fixes a third faster. Boris Cherny of Anthropic showed how Sonnet 5 and Sonnet 5.5 fix one project with failing tests. The new model finished in 12.3 seconds; the old one took 17.5.

A planet for a fifth of the price. Timothée Le Borgne, founder of OhMyUnicorn, gave Sonnet 5.5 and Opus 5.5 one prompt: a cartoon planet with cities, clouds and planes. Both built a living world, but Sonnet cost $0.96 against $4.79 for Opus.

Black holes at half the price. Independent benchmark BridgeBench gave Sonnet 5.5, Opus 5.5 and Fable 5.1 one prompt: an animation of two black holes merging. Sonnet 5.5 finished fastest and cost $0.34 against $0.69 for Opus 5.5.

A zombie shooter from one prompt. The BridgeMind team asked Sonnet 5.5 at maximum effort to build a 3D shooter. The game came out with a map, missions and reloading, but cost $177 and 49 minutes.

Financial documents. Box, an Anthropic partner, ran Sonnet 5.5 through its hard tasks with enterprise files. On financial documents, accuracy rose from 63% to 81% compared with Sonnet 5.

Our run. All of them passed, while Sonnet 5.5 cost $0.033 against $0.059 for Opus 5.5.

The difference shows in the details. In the tariff task, Sonnet 5.5 worked out a workaround on its own, which made it more expensive than Sonnet 5.

Where it stumbled. Tim Jayas gave Sonnet 5.5 and Fable 5.1 one prompt: a 3D model of a hummingbird micro-drone. Fable built the mechanics, labels and control panels, while Sonnet stopped at a single model, though it cost five times less.

More. Base44 got apps at the level of Opus 5 from Sonnet 5.5 in half the number of attempts, and blogger Jazii got a site five times cheaper than GPT-6 Astra. Simon Willison got 128K reasoning tokens in 15 minutes on max effort and never saw the image, while in FrontierCode this mode even cut the bill.

In the Claude app, Sonnet 5.5 is available on all plans, including the free one. Switch to it in the model list.

In Claude Code. Update the client with claude update: you need version 2.1.284 or later. Then /model sonnet in an open session.

Claude Code defaults to Opus 5.5, so you select Sonnet manually. Its default effort is medium.

Via API. The exact ID is claude-sonnet-5-5; on Bedrock, anthropic.claude-sonnet-5-5. The disabled reasoning mode is replaced by between_tools, and forcing a tool call returns a 400 error.

In clouds. On Bedrock, Google Cloud and Foundry, the short name sonnet in Claude Code still points to older versions. Use the full model name there.

In Anthropic's table, Sonnet 5.5 nearly caught up with Opus 5.5 and pulled far ahead of Sonnet 5. On Terminal-Bench 4.0 it even beat Opus's best result.

Independent index. Artificial Analysis ranked Sonnet 5.5 second: 56 points at maximum effort, while Opus 5.5 has 58. Up 18 points from Sonnet 5; the measurement was taken on a pre-release version.

In its own Terminal-Bench run, the lab also has Sonnet 5.5 ahead: 63.6% vs 59.6% for Opus 5.5. It trails in factual knowledge, with 54% accuracy against 66%.

Our paired API measurement against Sonnet 5 and Opus 5.5 is in chapter 3.

Rates are unchanged: $2 per million input tokens and $10 per million output tokens, same as Sonnet 5. That is exactly half the price of Opus 5.5.

Task cost. Anthropic promises up to 30% savings per task: the model spends fewer tokens on the same work. At Artificial Analysis, a task at medium effort costs $0.59, and at high effort $1.08.

At maximum effort the picture reverses. There, the task costs $7.60, one and a half times the cost on Sonnet 5.

Subscription. There is no separate subscription: Sonnet is included in all Claude plans, from Free to Max. Pro costs $20 a month, and the API is billed separately per token.

Sonnet 5.5 is the first Sonnet with cyber safeguards at the level of the top models. Risky security requests will be handled by Sonnet 5 instead, and the response will show that.

Injected commands. Sonnet 5.5 is more resistant than Sonnet 5 to commands hidden in files and web pages. In a Gray Swan test with 15 attempts, attacks succeeded in 3.4% of cases versus 6.7% for Sonnet 5.

In the browser, none of the attacks succeeded across 110 scenarios. Anthropic calls it the most robust model in the Sonnet class.

False alarms. In the first hours, Reddit users complained that the safeguards fired on ordinary code. Even a license-checking module got refused.

Honesty. According to the system card, the model misleads less often than Sonnet 5 and holds up more honestly under pressure than Opus 5.5. Its reasoning is harder to read than that of other recent Claude models.

Our experience. We build the site with agents on Fable 5.1 and Opus 5.5. We tested Sonnet 5.5 via the API on the night of release; results are in chapter 3.

Sonnet 5.5 or Opus 5.5. Anthropic recommends Sonnet for well-defined tasks and Opus 5.5 for complex work that requires judgment. Their rates differ by a factor of two.

On the Artificial Analysis index, the gap is two points. But at maximum effort, Sonnet writes more tokens, so its task costs more.

Pairing two models. Kevin Ngo of Creator says Sonnet 5.5 could be trusted to build a game from an architecture set by Opus 5.5. Anthropic also suggests Sonnet as a cheap complement to Opus at low effort.

Sonnet 5.5 or Sonnet 5. The price is the same, and the new model leads on every line of the announcement. The main reason to stay on Sonnet 5 in existing code: a forced tool call on 5.5 returns an error.

To choose for yourself, give both models one task you know and compare time, edits and the cost of the accepted result. Our own run is in chapter 3.

Sonnet 5.5 came out three months after Sonnet 5 and six days after Opus 5.5. Nine days later, Haiku 5.5 completed the family.