Anthropic releases Claude Sonnet 5.5 at the same list price as Sonnet 5

Same rate card, a different job

On 28 September 2026, Anthropic released Claude Sonnet 5.5, the second model in the Claude 5.5 family. It calls Sonnet 5.5 a faster, lower-cost complement to Claude Opus 5.5. In that split, Opus 5.5 is for complex work that needs careful judgment. Sonnet 5.5 is for well-scoped everyday tasks: fixing bugs, and writing documents, slides, and spreadsheets. Claude Haiku 5.5 is not in this release. Anthropic says it will follow in the coming weeks, with no date.

The list price matches Sonnet 5:

• $2 per million input tokens
• $10 per million output tokens
• $0.20 per million tokens for cache reads
• $2.50 per million tokens for cache writes

Opus 5.5, on the same table, is $4 input, $20 output, $0.20 cache reads, and $5 cache writes. The savings claim is about tokens used, not the rate. Anthropic says that in its own testing a task costs up to 30 percent less than on Sonnet 5, and that output is more than 30 percent faster.

The default effort is Medium in Claude Code and Anthropic’s apps, and High on the Claude Platform. On the company’s own cost charts, Sonnet 5.5 at low or medium effort beats Sonnet 5’s best score at roughly a tenth of the cost per task.

Scores Anthropic published

Anthropic says Opus 5.5 remains clearly stronger at complex, open-ended work. Figures in the announcement include:

• Terminal-Bench 4.0: Sonnet 5.5 70.6 percent, Sonnet 5 10.3 percent, Opus 5.5 66.4 percent at Xhigh, which Anthropic says is that model’s highest score. It does not report GPT-6 Sol, because it says neither Terminal-Bench nor OpenAI published one.
• CursorBench 4.0: 55.5 percent, versus 34.1 percent for Sonnet 5 and 57.8 percent for Opus 5.5.
• FrontierCode 1.1 (main), scored on whether a change would be merged: 46.2 percent at Max and 52.1 percent at Xhigh, versus 42.4 for Sonnet 5, 54.4 for Opus 5.5, and 49.3 for GPT-6 Sol. Max came in below Xhigh. Anthropic says that at Max the model more often ran a multi-agent code review, and that in two cases Cognition looked at, that caused a timeout or out-of-scope edits, which this test penalizes.
• GDPval-AA v2.1: 1844, versus 1449 for Sonnet 5, 1846 for Opus 5.5, and 1487 for GPT-6 Sol.
• OSWorld 2.1 (partial): 80.1 percent, versus 57.0 and 81.8. Humanity’s Last Exam, with tools: 64.5 percent, versus 54.9 and 67.7.

Artificial Analysis ran GDPval-AA and a related briefcase test on a pre-release deployment Anthropic says had a structured-output bug, since fixed. Anthropic says any effect likely understated Sonnet 5.5. It also says OpenAI has fixed an image-understanding bug in GPT-6 Sol, so some third-party comparison scores may not yet match the current model. Customer comments in the post are testimonials, not public datasets. Slack’s Curtis Allen said that, with prompts unchanged, Sonnet 5.5 beat Sonnet 5 on almost all offline Slackbot evals, in fewer steps and with about 14 percent fewer output tokens.

Controls, and how to call it

Anthropic says Sonnet 5.5 does not set a new capability frontier for the company. On an automated audit of roughly 1,850 scenarios, it says the model improves on or matches Sonnet 5 on most alignment measures, and that no suite catches every failure.

Because it rates cybersecurity capability as comparable to Opus 5, this is the first Sonnet with cyber safeguards and fallbacks like those on its strongest models. Routine bug-fixing still runs. Higher-risk cybersecurity requests fall back to Sonnet 5. A Cyber Verification Program for tiered access is described as coming soon. Biology safeguards are unchanged from Sonnet 5. Sonnet 5.5 also adds classifiers aimed at blocking extraction of reasoning, and expanded “preserved thinking,” so a chain of thought stays tied to the account that created it.

The model is offered with zero data retention, on Amazon Web Services, Google Cloud, and Microsoft Azure, and on the Claude Platform as `claude-sonnet-5-5`. Requests that turn thinking off must switch to a `between_tools` setting first. The scores above are Anthropic’s launch figures, with methods described in the system card the announcement points to.