Anthropic's Claude Sonnet 5.5 is faster and cheaper per task, and beats its flagship on one benchmark

The model most apps actually run on

Anthropic released Claude Sonnet 5.5 on 28 September 2026, an update to the mid-tier model that sits between the small, cheap Haiku and the flagship Opus. Sonnet is the one that matters most in practice, because it is the model most coding tools, chatbots and business apps default to when they need real capability without the top-tier bill. This release follows Opus 5.5, which shipped on 22 September, and delivers on Anthropic's promise to refresh Sonnet in the coming weeks.

The pitch is simple and, unusually for a model launch, mostly about economics rather than raw intelligence. Sonnet 5.5, in Anthropic's words, "runs more than 30% faster, and costs up to 30% less for most work" than the Sonnet 5 it replaces. It is now live in Claude Code and across the major clouds.

What actually changed: speed and cost

Start with the number that is easy to misread. The headline price did not fall. Sonnet 5.5 keeps Sonnet 5's pricing exactly: $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 and cache writes at $2.50 per million. Sonnet 5 has carried that same $2 and $10 price since it launched on 30 June 2026.

So where does "up to 30% less" come from? From efficiency, not a discount. A faster model that reaches the same answer in fewer output tokens costs less to run per task even at an identical per-token rate, and output tokens are the expensive half of the bill. That is a real saving, but it is an effective one that shows up in your usage, not a lower sticker you can point to on the pricing page. For why output tokens dominate an AI bill, see our guide to how AI pricing works.

The benchmark that turns heads

The eye-catching claim is that a mid-tier model can, on some tasks, out-score the flagship above it. On Terminal-Bench 4.0, an agentic test of a model driving a command line, Anthropic reports Sonnet 5.5 at 70.6%, ahead of Opus 5.5 at 66.4% and far ahead of the old Sonnet 5 at 10.3%. On OSWorld 2.1, a computer-use benchmark, it scores 80.1%, a whisker behind Opus at 81.8%. On Humanity's Last Exam with tools it reaches 64.5%, close to Opus's 67.7%.

The picture is not a clean sweep, and that is the honest part. On FrontierCode 1.1, a coding benchmark, Opus 5.5 still leads clearly, 54.4% to Sonnet's 46.2%. On CursorBench 4.0 Opus is ahead too, 57.8% to 55.5%. The pattern is that Sonnet 5.5 has closed most of the gap to the flagship on agent and computer-use tasks while still trailing on the hardest raw coding. For most everyday work, a model that matches or nearly matches Opus at Sonnet's price is the story. For the frontier of difficulty, Opus is still the flagship.

The catch: every number here is vendor-reported

There is one caveat that applies to the entire benchmark table, and it is worth stating plainly. All of these scores come from Anthropic, testing its own models. That is normal for a launch, and Anthropic's benchmark reporting is generally careful, but a company grading its own homework is not the same as an independent evaluation. A figure like Sonnet 5 scoring 10.3% on Terminal-Bench 4.0 also signals that this is a newer, harder version of the test, which makes the comparisons within the table meaningful but comparisons to older reported scores misleading. Independent evaluations from groups that run their own harnesses have not landed yet. Until they do, treat the leaderboard as the vendor's claim, and the speed and cost-per-task improvements, which you can measure yourself, as the firmer ground.

A first for Sonnet on safety

One genuinely new thing sits underneath the performance numbers. Sonnet 5.5 is the first Sonnet model to launch with cyber safeguards and fallbacks similar to those Anthropic uses for its most capable models, with cybersecurity capabilities it describes as comparable to Opus 5. Its biology safeguards are unchanged from Sonnet 5. Both, Anthropic says, target a narrow set of high-risk requests and leave routine software development and most life sciences work untouched. It is a sign that as the mid-tier model grows more capable, guardrails once reserved for the flagship move down the range with it.

Where you can use it

Sonnet 5.5 is available now through the Claude platform, Amazon Web Services, Google Cloud and Microsoft Azure, and it is already the default in Claude Code. The API model string is claude-sonnet-5-5, and zero data retention is available for teams that need it. If you are weighing Claude against the other big assistant, our Claude versus ChatGPT comparison walks through how the two families differ.

Our take

This is a quietly important release precisely because it is not a flashy one. There is no new headline capability, no price war, no benchmark that redraws the map. What Anthropic has done is make the workhorse model faster and cheaper to actually run, and pull it close enough to the flagship on agent tasks that many teams paying for Opus will ask whether they still need to. That is the kind of change that shows up in bills and product decisions rather than headlines. The two claims to trust are the ones you can verify: it is faster, and it costs less per task at the same rate. The claim to hold loosely is the near-flagship framing, not because it is implausible, but because the only scorecard so far was written by the team that built the model.