Anthropic Releases Claude Sonnet 5.5, Now Powering Free Tier on Claude.ai

Anthropic Releases Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026, positioning it as a direct upgrade to Sonnet 5. According to Anthropic, the new model "runs 30%+ faster, and costs up to 30% less for most work" while matching Sonnet 5's list price. Independent testing by developer Simon Willison found Sonnet 5.5 appears to beat its predecessor on every benchmark he checked, though Anthropic has not published a full benchmark table alongside the release.

Pricing has not been disclosed in exact per-token terms; Anthropic states the sticker price matches Sonnet 5, with actual costs coming down due to efficiency gains rather than a rate cut. Context window size and training cutoff date were not specified in the announcement.

Thinking mode bug carries over from Opus 5.5

Sonnet 5.5 ships with adjustable "thinking effort" settings, including a "max" tier that reportedly inherits a bug previously observed in Claude Opus 5.5: at max effort, the model can burn through 128,000 tokens (costing $1.28 in one test) without producing a final output. At the "xhigh" thinking setting, the same test — generating an SVG of a pelican riding a bicycle — completed in 41 seconds at a cost of 5.74 cents and produced a structurally correct result, complete with an accurate bicycle frame and pedal placement, though the pelican's helmet was described as "misshapen."

On coding tasks, including widely-shared 3D animation demos, Sonnet 5.5 reportedly performs close to Claude Opus 5.5's level, according to early testing shared on social media.

Now the default free-tier model

The most consequential change may be Sonnet 5.5's deployment as the model behind the free tier of claude.ai. This puts it in direct comparison with OpenAI's free ChatGPT tier, which currently runs on a model referred to as "Luna 5.6." Based on hands-on testing — including a prompt asking for an HTML page rendering a 3D WebGL pelican on a bicycle, which produced a functional result — Anthropic's free tier now appears meaningfully more capable than OpenAI's equivalent offering.

Anthropic also reiterated in its announcement that a smaller Haiku 5.5 model is coming "in the coming weeks," without providing a firm release date or pricing.

What this means

Anthropic is using efficiency gains, not price hikes, to compete: same list price, better benchmarks, and lower effective compute cost per task. Putting Sonnet 5.5 behind the free tier is a direct swipe at OpenAI's free ChatGPT offering, betting that a stronger free experience drives conversion and mindshare even without extracting revenue from that tier. The recurring thinking-mode bug — where maximum reasoning effort burns tokens without producing output — is now a two-release pattern (Opus 5.5 and Sonnet 5.5), suggesting a shared architectural issue in Anthropic's extended-thinking implementation that will need a fix before enterprises can trust the "max" setting for production workloads. The real test will be Haiku 5.5's pricing, which Anthropic has not yet disclosed but which will determine whether it can compete with OpenAI's cheaper Luna-tier models on cost-sensitive, high-volume use cases.