What is Claude Sonnet 5.5, and how is it faster and cheaper to run?

Yesterday, Monday, September 28, 2026, Anthropic released Claude Sonnet 5.5, the second member of its Claude 5.5 family after Claude Opus 5.5, which arrived less than a week earlier. Sonnet has long been Anthropic's middle model: less powerful than Opus but quicker and cheaper. TechCrunch sums up this release in one word: speed.

Anthropic positions the two models as a team. Opus 5.5 is for complex work that needs careful judgment. The new Sonnet shines on well-defined everyday jobs: fixing bugs, writing, and producing polished documents, slides and spreadsheets. A third, smaller model, Claude Haiku 5.5, is promised in the coming weeks.

What Anthropic released

Sonnet 5.5 is available today in the Claude apps, in Claude Code, on the Claude Platform for developers (as claude-sonnet-5-5), and on the AWS, Google Cloud and Azure cloud platforms.

The headline changes compared with Sonnet 5:

Speed: output arrives more than 30% faster, making it Anthropic's fastest Sonnet so far.
Cost per task: up to 30% lower, even though the price per token is unchanged.
Ability: large gains in coding and office work. On Terminal-Bench 4.0, a test of multi-step tasks in a command line, Anthropic reports 70.6%, against 10.3% for Sonnet 5.
Writing: like Opus 5.5, it writes more clearly, and early testers called it a better collaborator.

As always, these figures are Anthropic's own. The company itself cautions that benchmark scores capture only one side of a model, and says Opus 5.5 remains clearly stronger on open-ended work that needs sustained judgment.

How it gets cheaper without a price cut

The price per token stays the same as Sonnet 5. A token is a small chunk of text, about part of a word, and you pay for tokens you send (input) and tokens the model writes (output). So where do the savings come from? Sonnet 5.5 simply uses fewer tokens for the same job. Anthropic says it groups tool calls together more often, so it finishes in fewer steps. Fewer steps mean less text read and written, which means a smaller bill.

Here is the price list Anthropic published at launch. Cache writes and cache reads refer to the prompt cache: when a program sends the same long instructions or documents again and again, the Claude Platform can store them once and let later requests "read" that stored copy for a fraction of the normal input price.

Per million tokens	Sonnet 5.5	Opus 5.5
Input	$2	$4
Output	$10	$20
Cache writes	$2.50	$5
Cache reads	$0.20	$0.20

A quick worked example shows why caching matters. Suppose a customer-support bot sends a 50,000-token help manual with every one of 1,000 questions a day. That is 50 million tokens of manual. Sent as ordinary input, it costs about $100 a day; read from the cache, about $10. And because Sonnet 5.5 typically needs fewer tokens to finish each job, the rest of the bill shrinks on top of that.

Effort, speed and quality

Like other recent Claude models, Sonnet 5.5 has an effort setting: how long it thinks and how carefully it checks itself before answering. The default is Medium in Claude Code and the Claude apps and High on the Claude Platform. Anthropic reports that at Low or Medium effort Sonnet 5.5 often tops Sonnet 5's best result while costing roughly one tenth as much per task, and that at Max effort it comes close to Opus 5.5 on several tests.

That makes it a natural "first try" model. You can let it work fast, then raise the effort or switch to Opus 5.5 only when a task proves hard.

Prompt

You are helping me fix a small bug. In the attached file, the "total"
function returns the wrong value when the cart is empty. Find the cause,
change as few lines as possible, and show me the diff with a one-line
explanation. Don't refactor anything else.

A tightly scoped request like this plays to Sonnet 5.5's strengths: a clear goal, a small change and a short answer.

Try it: Give Sonnet 5.5 the same short task twice, once at Low effort and once at High, such as summarizing a two-page article in five bullet points. Compare the answers and the time each takes, and decide which level you need for that kind of job.

New safeguards for a Sonnet model

Sonnet 5.5's cybersecurity skills are, by Anthropic's account, comparable to Opus 5's. For the first time, a Sonnet model therefore launches with the cyber safeguards Anthropic uses on its most capable models. Routine work such as finding and fixing bugs in your own code is unaffected, but higher-risk security requests visibly fall back to Sonnet 5. Its biology safeguards stay the same as Sonnet 5's.

It is also the first Sonnet with protection against distillation: attempts to copy a model's abilities by harvesting huge numbers of its answers through fake accounts, then training another model on them. Classifiers now block efforts to extract its reasoning at scale.

Why it matters

The AI race is increasingly about cost and speed, not only peak intelligence. Businesses running thousands of tasks a day care as much about the bill and the waiting time as about the top score. TechCrunch notes that Sonnet 5 was pitched three months ago on running agents cheaply; Sonnet 5.5 pushes the same idea further. It also arrives in a busy month, a week after OpenAI updated its own mid-tier and budget models.

How do you get the most out of Claude Sonnet 5.5's speed and price?

Use it as your default for well-defined work: bug fixes, drafting, summaries, data cleanup and document or slide creation. Keep the effort at Medium or Low for routine requests and raise it only when an answer falls short. Save Opus 5.5 for open-ended problems that need long, careful judgment.

If you build with the API, structure your prompts so the unchanging part, such as instructions, policies or reference documents, comes first and can be cached. Because a cache read costs a tenth of the normal input price, a long, stable prefix becomes cheap to resend. Finally, measure on your own tasks: compare cost, time and quality against the model you use today before switching everything over.