Anthropic has released Claude Sonnet 5.5, promising a model that is more than 30% faster than Sonnet 5 and can cost up to 30% less per task without raising its headline token price. The pitch is unusually practical: not the company’s most powerful model, but a faster everyday workhorse for coding, documents, slides, spreadsheets and other well-scoped jobs.
That distinction matters in a market where model launches can be reduced to benchmark scoreboards. Anthropic still positions Opus 5.5 as the stronger choice for complex, open-ended work requiring sustained judgement. Sonnet 5.5 is being sold on something businesses feel more directly: how quickly useful work gets done and what that work costs.
Claude Sonnet 5.5 keeps the token price but uses fewer of them
Anthropic lists Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, the same headline prices as Sonnet 5. The claimed saving comes from efficiency. The company says the new model typically needs fewer tokens to complete the same task, which can reduce the total bill even though the unit price has not changed.
The company’s main Sonnet product page also presents the model as the default choice for a wide range of professional work. That is a useful clue to the commercial battle now taking shape. AI providers are no longer competing only over the hardest reasoning problem a model can solve. They are competing over how much capable work a customer can push through a model every hour and every pound or dollar.
Speed is part of that equation. Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5. For an individual chat that may mean a shorter wait. In software development or agent systems, where one task can trigger many sequential calls, the same improvement can compound across a workflow.
The coding jump is large, but it is still vendor-reported
Anthropic highlights a striking result on Terminal-Bench 4.0, an evaluation of multi-step command-line tasks. It reports Sonnet 5.5 at 70.6%, compared with 10.3% for Sonnet 5. The company also reports gains on coding, computer-use and knowledge-work evaluations, and says the new Sonnet can in some settings approach Opus 5.5.
Those figures deserve attention, but they should not be treated as a universal measure of how much better the model will feel. Benchmarks are controlled tests, model settings matter, and the vendor is reporting its own launch results. Anthropic itself notes that Opus remains clearly stronger in its testing for complex, open-ended work requiring sustained judgement.
That gap between benchmark success and everyday reliability is familiar. LiveAIWire has previously examined Anthropic’s work on AI agents operating laboratory equipment, where useful performance depends on sustained action and the consequences of mistakes, not simply a one-shot answer. Faster models become more valuable as they take on longer workflows, but reliability becomes more important at the same time.
Sonnet is moving closer to the frontier safety tier
Anthropic says Sonnet 5.5 is the first Sonnet model launched with cyber safeguards and fallbacks similar to those it developed for its most capable models. The company says this reflects cybersecurity capabilities comparable to Opus 5, while biology safeguards remain the same as Sonnet 5.
The company also says an automated behavioural audit covering roughly 1,850 scenarios found Sonnet 5.5 improved on or matched Sonnet 5 on most measures of alignment, resistance to misuse and honesty. As with performance benchmarks, these are Anthropic’s own evaluations. The value is in showing which risks the developer believes have become important enough to require stronger controls in a mid-tier model.
That is a notable shift. The more capable the everyday model becomes, the less useful it is to think of safety as something reserved for a single premium frontier system. A model intended for high-volume routine work can touch more code, files and business processes precisely because it is cheaper and faster to deploy.
The real contest is now cost per completed job
For users choosing between models, the headline token rate is increasingly an incomplete comparison. Two models can charge the same per token while producing very different total bills if one needs fewer tokens, fewer retries or less human correction. The same is true for speed: a faster first response is useful, but a faster reliable completion is what matters.
The release also comes as Anthropic is under pressure from more than technical rivals. LiveAIWire has covered the company’s legal dispute with major music publishers, a reminder that the economics of frontier AI sit alongside arguments about training data, rights and responsibility. Model efficiency does not make those questions disappear.
Claude Sonnet 5.5 is therefore an important launch even if it never tops every leaderboard. It shows where the market is heading: strong models pushed down into the everyday tier, where latency, total task cost and dependable execution matter as much as a spectacular benchmark. For most organisations, that is where AI will either become ordinary infrastructure or remain an expensive experiment.
For buyers, that creates a more useful evaluation question than “which model wins?” A support team may care about response time and consistency, a developer about code changes that survive testing, and a finance team about the full cost of thousands of repeated runs. Sonnet 5.5 is designed to compete on that operational layer. If Anthropic’s efficiency claims hold outside its own tests, the model could be more consequential as a volume tool than as a benchmark headline.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.