Claude Sonnet 5.5: near-Opus results, 30% faster, same price

Anthropic released Claude Sonnet 5.5 on September 28, 2026, less than a week after Claude Opus 5.5. It is the second model in the Claude 5.5 family and the new default Sonnet. The price per token hasn't changed. It is faster, it uses fewer tokens, and on many tests it comes close to Opus 5.5.

What is new in Sonnet 5.5

Anthropic describes Sonnet 5.5 as "a clear upgrade over Claude Sonnet 5". It is meant to sit alongside Opus 5.5 as the faster, cheaper choice for clearly defined work.

• Everyday work, done well. It is strongest at well-scoped coding tasks, bug fixing, and polished documents, slides and spreadsheets.
• Fewer steps to finish a job. Like Opus 5.5, it usually needs far fewer requests and tool calls than its predecessor to finish an agentic task. Most of the savings come from that.
• Clearer writing. Early testers said it was a better partner to work with than Sonnet 5, mostly because it writes more clearly. In one internal test, two experts judged its first draft of a 10-slide operating review ready to send as it was.
• Better at using a computer. It is the first Sonnet model to beat Pokémon Red working only from screenshots. On OSWorld 2.1 it scores 80.1%, up from 57.0% for Sonnet 5.
• The same big context. It keeps the 1 million token context window and 128K token maximum output.

Pricing: same price, smaller bills

The per-token price is the same as Sonnet 5's. Here is how it compares with Opus 5.5:

Batch processing still saves 50%, and US-only inference costs 1.1 times the standard price.

So where does "up to 30% cheaper" come from? From using fewer tokens. In Anthropic's tests, Sonnet 5.5 often matched Sonnet 5's best score at low or medium effort for a fraction of the cost:

• On FrontierCode, at high effort it scored 10 points higher than Sonnet 5 at about one fifteenth of the cost per task.
• On several benchmarks, at low or medium effort, it beat Sonnet 5's best result for about a tenth of the cost.

Customers report similar gains. Slack measured about 14% fewer output tokens than Sonnet 5 in its Slackbot evaluations. Balyasny Asset Management saw answers to finance tasks drop from about 497,000 tokens to 121,000. Lovable reported a third fewer tool calls in its coding evaluations.

Benchmarks

Here is how Sonnet 5.5 compares on a selection of the benchmarks Anthropic published:

The jump from Sonnet 5 is huge on almost every row. More interesting is how close Sonnet 5.5 gets to Opus 5.5, which costs twice as much per token. It is level on GDPval-AA knowledge work and within about two points on CursorBench. On Terminal-Bench 4.0 it even scores higher.

Anthropic is careful not to oversell this. At maximum effort, Sonnet 5.5 performs about as well as Opus 5.5 on several tests, but Opus 5.5 remains "clearly stronger at complex, open-ended work".

As always, benchmarks are a guide, not a guarantee. Test the model on your own tasks before you commit.

Safety

On Anthropic's automated behavioural audit, which covers about 1,850 scenarios, Sonnet 5.5 matches or improves on Sonnet 5 on most measures. On tests of whether it tries to escape its sandbox, it comes close to Opus 5.5. Anthropic says it found no evidence that the model pursues goals that conflict with what users want.

There are also new safeguards:

• Cybersecurity: it is the first Sonnet model to launch with the same kind of cyber safeguards used on Anthropic's top models. Routine bug fixing is not affected. Higher-risk cyber requests visibly fall back to Sonnet 5. Security professionals can apply to the Cyber Verification Program for wider access.
• Biology: the safeguards are the same as on Sonnet 5. Some microbiology and virology requests may be flagged by mistake, and researchers can apply to the Life Sciences Verification Program.
• Reasoning extraction: it is the first Sonnet with classifiers that stop people from extracting its internal reasoning, for example to train a competing model.

Where you can use it

• Claude apps: anyone can chat with Sonnet 5.5 on Claude.ai on the web, iOS and Android.
• Claude API: use the model ID claude-sonnet-5-5.
• Cloud platforms: Amazon Bedrock (anthropic.claude-sonnet-5-5), Claude Platform on AWS, Google Cloud and Microsoft Foundry.

Zero data retention is available. In Claude Code and the Claude apps, the default effort level is medium. On the API it is high.

For developers: what changes when you upgrade

Most Sonnet 5 prompts work well on Sonnet 5.5 straight away. The API has five breaking changes, though. Four of them match the ones that came with Opus 5.5.

Turning thinking off: meet between_tools

On Sonnet 5 you could switch thinking off. On Sonnet 5.5 the lowest setting is thinking: { type: "between_tools" }. The model skips extended thinking, but still writes short progress notes between tool calls. It has a few limits: it only works at high effort or below, you can't add other fields to the thinking object, and no other model accepts it.

Anthropic recommends trying adaptive thinking at low effort first. At that level the model skips thinking on most simple requests anyway.

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic();

// Recommended: keep adaptive thinking on and set the effort level yourself.
const message = await client.messages.create({
  model: 'claude-sonnet-5-5',
  max_tokens: 16000,
  output_config: { effort: 'medium' }, // low | medium | high | xhigh | max
  messages: [{ role: 'user', content: 'Fix the failing test in utils.test.ts.' }],
});

// If a route must stay thinking-off, use between_tools at high effort or below.
const quick = await client.messages.create({
  model: 'claude-sonnet-5-5',
  max_tokens: 16000,
  thinking: { type: 'between_tools' },
  output_config: { effort: 'high' },
  messages: [{ role: 'user', content: 'Summarise this changelog in three bullets.' }],
});

Effort levels have been recalibrated

This one is easy to miss because nothing breaks. The default is still high, but each level no longer produces the same amount of thinking as on Sonnet 5. Don't just carry your old setting over. Re-test, and start with medium for agentic coding and multi-step tool use, and low for chat, classification and extraction.

Progress notes have moved

During long agent runs, the short notes the model writes between tool calls now arrive as thinking blocks, which are empty by default. If your app shows those notes to users, it may look like it has gone silent. Set thinking: { type: "adaptive", display: "updates" } (currently a beta feature) to get them back as short summaries. With between_tools, the notes come back as readable text automatically.

New features Sonnet 5 didn't have

Sonnet 5.5 also picks up several features that were missing from Sonnet 5: system messages in the middle of a conversation, changing effort per message (beta), task budgets (beta) and mid-conversation tool changes (beta). The minimum prompt length for caching also drops from 1,024 to 512 tokens. Note that Sonnet 5.5 has its own rate limits, separate from Sonnet 5's, so check your tier before you move a lot of traffic.

Sonnet 5.5 or Opus 5.5?

With the two models this close, the choice comes down to the kind of work:

• Pick Sonnet 5.5 for well-defined tasks where speed and cost matter: bug fixes, routine coding, support tickets, documents, slides and spreadsheets, and agents that run at high volume.
• Pick Opus 5.5 for long, open-ended work: big refactors, hard research questions and long-running agents where getting it right first time is worth twice the price.
• Still on Sonnet 5? Switch. You get a faster model that scores higher on every benchmark Anthropic published, for the same price per token. Just fix the breaking changes above and re-test your effort level.

What comes next

Anthropic says Claude Haiku 5.5 will join the Claude 5.5 family "in the coming weeks", built for high-volume, cost-sensitive work. That would complete the line-up: Haiku for scale, Sonnet for everyday work and Opus for the hardest jobs. We will cover it when it lands.