Claude Sonnet 5.5

What it claims

Anthropic released Claude Sonnet 5.5 on 28 September 2026. It is priced the same as Sonnet 5, and Anthropic says it "runs 30%+ faster, and costs up to 30% less for most work". Willison says it appears to beat Sonnet 5 on every benchmark, though the post cites no benchmark names or scores. Willison also judges it "almost as good as Opus 5.5 on some coding tasks", including viral 3D animation demos. Sonnet 5.5 is now the model behind the claude.ai free tier. ChatGPT's free tier runs Luna 5.6, and Willison calls Anthropic's free offering "much more capable".

Evidence

The post contains no benchmark tables, only Willison's informal pelican-on-a-bicycle SVG test at different thinking-effort settings:

On the free tier, the prompt build me an HTML page that renders a three-dimensional pelican riding a bicycle using WebGL returned what Willison calls "a solid effort". This is a single qualitative sample.

How to use it

• The thinking-effort levels named in the post include max and xhigh.
• On this test, xhigh cost about 1/22 of what max cost and actually produced output.
• That makes xhigh the practical ceiling to start from until the max behaviour below is fixed.

Limits

• Runaway thinking at max effort: this is the same bug Willison saw with Opus 5.5. The model spent its whole 128,000-token budget on thinking and returned nothing, and the request was still billed ($1.28). If you route requests to max, set token caps and handle empty responses.
• No measured claims: the speed and cost claims are Anthropic's figures, and the post includes no measurements of its own.
• Coding comparison: "almost as good as Opus" is an impression from a few demos, not a measurement.
• Haiku 5.5: Anthropic says it is coming "in the coming weeks". Willison hopes it will be price-competitive with GPT-6 Luna. Its pricing is not yet known.