Claude Sonnet 5.5: what has really changed since launch

Anthropic released Claude Sonnet 5.5 on 28 September 2026, six days after Opus 5.5. Since then, my feed hasn’t talked about anything else. I read the announcement, the docs and a good part of the feedback to find out what this model actually changes when you code with an AI.

Faster, and cheaper where it counts

The price per token hasn’t moved: $2 per million input tokens, $10 for output, the same as Sonnet 5. What changes is how many tokens a job takes to finish. Anthropic announces a model more than 30% faster and up to 30% cheaper per task.

The testimonials published with the announcement point the same way. Slack reports about 14% fewer output tokens. Box measures a model 2.4 times faster, with 12% fewer tokens. These are figures from customers chosen by the publisher, so treat them as a trend rather than a guarantee.

The benchmarks, with the usual caveats

The most dramatic jump concerns terminal work: 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5. On GDPval-AA, which measures professional tasks, it is almost level with Opus 5.5 (1,844 against 1,846). Opus keeps the edge on seven of the eight benchmarks in the official comparison, by three points at most on those scored as percentages.

In other words, Sonnet 5.5 costs half as much as Opus 5.5 ($4 and $20 per million tokens for the latter) and gets close to it on everyday work. Cursor recommends it, moreover, since it scores higher on CursorBench at the same price per token.

What the API adds

The model chooses for itself how much to think, with five effort levels: low, medium, high, xhigh and max. Developers also gain a few options:

• change the effort level message by message without losing the prompt cache
• insert system messages in the middle of a conversation
• declare tools partway through an exchange
• request context compaction on demand

Nothing spectacular taken alone. Together, they make long agent sessions cheaper to run.

On safety

It’s the first Sonnet shipped with cyber safeguards of the same kind as Opus’s. Anthropic adds classifiers to prevent extraction of the model’s reasoning, a defence against distillation.

What doesn’t work yet

Enthusiasm isn’t unanimous. Simon Willison reported that at max effort, the model burned 128,000 tokens of reasoning in fifteen minutes without producing his SVG, the same failure as with Opus 5.5. An early-access user, Alex Finn, describes a subtle improvement, noticeable mainly after a few days of use, not a leap. The highest effort level isn’t always the right setting.

The wave around it

Sonnet 5.5 landed right in the middle of a noisy wave: the game “mashup” videos, where Minecraft runs inside Elden Ring. They’re driven by Opus 5.5 rather than Sonnet, but they show what these models can do on their own when they’re handed whole programs. I’m dedicating another article to them.

My reading stays modest: Sonnet 5.5 is the reasonable default model for code, and Opus is for the cases where it gets stuck.