Claude Sonnet 5.5: Anthropic lifts its mid-range model to top-tier level, at half the price

Just six days after unveiling Claude Opus 5.5, its most powerful model, Anthropic released the second member of this new generation on 28 September 2026: Claude Sonnet 5.5. In the American company’s line-up, Sonnet sits in the middle, between Haiku, small and cheap, and Opus, the most capable and most expensive. It is the model Anthropic presents as the everyday workhorse, meant to offer the best balance between quality and cost.

Anthropic’s message boils down to three promises: a faster model, a lighter bill and a level close to the top of the range. The tests published by the company point in that direction, with one major caveat: they are its own measurements. The first independent evaluations, and above all real-world use, will show whether these gains hold up.

Same price, lower bill

Anthropic has not changed the list price. On its developer platform, Sonnet 5.5 costs $2 per million tokens sent to the model and $10 per million tokens generated in response, exactly like Sonnet 5. Tokens are fragments of words used as a unit of account: a page of text represents a few hundred of them. Opus 5.5 costs twice as much, $4 and $20.

If the bill goes down, it is because the model works more economically. According to Anthropic, it uses significantly fewer tokens to complete the same task, and more often bundles several actions into a single step. The announced result: a cost per task up to 30% lower than Sonnet 5, and responses generated more than 30% faster. At Slack, one of the clients cited by the company, the model outperformed its predecessor on most internal tests of the in-house assistant, while writing about 14% fewer tokens, without the prompts being changed.

• Sonnet 5.5: $2 per million input tokens, $10 output (unchanged).
• Opus 5.5: $4 and $20, twice as much.
• Reusing context already sent (caching): $0.20 per million tokens for both models.
• Gains announced by Anthropic: over 30% faster, up to 30% cheaper per task.

A spectacular leap in coding

The most striking progress concerns autonomous code writing. On Terminal-Bench 4.0, a test in which the model must carry out multi-step tasks on its own in the terminal, the command screen used by programmers, Sonnet 5 passed 10.3% of the exercises. Its successor passes 70.6%, and even beats Opus 5.5, credited with 66.4%. On CursorBench 4.0, built from real sessions of the Cursor code editor, it rises from 34.1% to 55.5%, just over two points behind Opus 5.5.

Not everything is uniform. On FrontierCode, a test that checks whether a code change would be accepted as-is in a real project, the figures vary with the level of reasoning requested from the model: 52.1% at the “Xhigh” setting, but 46.2% at the maximum setting, against 54.4% for Opus 5.5. Anthropic attributes this to the model’s tendency, at that setting, to launch code reviews by several sub-agents more often, which sometimes led it to run out of time or to change parts it had not been asked to touch.

Almost Opus for office work

On GDPval-AA, which measures the quality of work on concrete tasks drawn from 44 occupations, Sonnet 5.5 scores 1,844 points, against 1,846 for Opus 5.5 and 1,449 for Sonnet 5: practically a tie with the top model. The same trend appears on OSWorld 2.1, which assesses the ability to use a computer like a human (80.1% against 81.8% for Opus 5.5). Anthropic acknowledges, however, that these figures do not tell the whole story: on long, loosely defined work that requires judgement, Opus 5.5 remains clearly superior, it says.

The division of roles is therefore deliberate. Opus 5.5 gets the complex missions; Sonnet 5.5 the everyday, well-defined tasks: fixing bugs, producing documents, presentations or spreadsheets. The client companies cited point the same way. Zendesk, the customer service specialist, says it handled its support requests 20% faster with fewer decision errors; Epic Games, the publisher of Fortnite, says the model reached the level expected of a higher-tier model on software design audits. Anthropic also slips in an anecdote: it is the first Sonnet to finish Pokémon Red, the Game Boy game, relying on screenshots alone.

Top-tier safeguards

This gain in power has a downside, which Anthropic highlights. In cybersecurity, Sonnet 5.5’s capabilities come close to those of Opus 5, the top model of the previous generation. They could therefore also be useful to hackers. For the first time, a Sonnet launches with protections comparable to those of the company’s most powerful models: requests deemed risky in cybersecurity are handed, visibly to the user, to Sonnet 5, which is much less capable on these topics. Finding and fixing ordinary bugs is not affected.

Another new feature: filters against “distillation”. This technique consists of querying a model massively, often through many accounts, to harvest its answers and train a competing model on the cheap. Labs legitimately use it on their own models to derive lighter versions; it becomes a problem when a competitor uses it without permission. Anthropic had already accused several Chinese labs of doing so. In biology, the protections remain those of Sonnet 5, designed to block dangerous requests without hindering research, teaching or medical work.

Against OpenAI, a tricky comparison

The competition is head-on. OpenAI launched GPT-6 Sol on 22 September, the same day as Opus 5.5, and charges exactly the same price as Sonnet 5.5. Comparisons remain fragile, however: OpenAI has not published a Terminal-Bench 4.0 score, and the measurement made by the independent firm Artificial Analysis, 43% at the maximum setting, was obtained with OpenAI’s coding tool and under a different protocol. Where the two models appear in the same tables, Sonnet 5.5 comes out ahead on GDPval-AA, with 1,844 points against 1,487, according to the figures published by Anthropic.

This race of announcements comes at a particular moment for Anthropic. According to 01net, the company is preparing a stock market listing, which could take place in October, and every new model also serves to win over investors. Its CEO, Dario Amodei, nonetheless keeps warning about the risks of AI, a paradox his critics are quick to point out.

Availability

Sonnet 5.5 has been available since 28 September in the Claude apps, on the web, iOS and Android, as well as on Anthropic’s developer platform and through the major cloud providers: Amazon Web Services, Google Cloud and Microsoft. In the apps and in Claude Code, it runs by default at a “medium” effort level, a setting that trades off speed against deeper reasoning; on the developer platform, the default setting is “high”. Finally, Anthropic announces that Claude Haiku 5.5, intended for very high-volume uses, will arrive in the coming weeks.

Beyond the figures, this launch confirms an underlying trend: mid-range models are rapidly catching up with the top models of the previous generation, and sometimes of the current one, for a fraction of the price. For businesses, the question is no longer just which model is the most powerful, but which offers the best balance of quality, speed and cost for each task. Benchmarks published by the vendors give a first indication; only tests in real conditions, on your own data, can settle it.

NovoTech uses artificial intelligence tools, including Anthropic’s Claude, to help with news monitoring. This analysis is based exclusively on Anthropic’s official announcement and on the public articles cited below.