Anthropic launches faster Sonnet 5.5 for everyday work
Anthropic released Claude Sonnet 5.5 on 28 September, positioning it as a faster, more economical default for coding and everyday professional work. It generates output more than 30% faster than Sonnet 5 and costs up to 30% less per task in Anthropic's tests while retaining the same $2 per million input tokens and $10 per million output tokens. The improvement comes from doing work with fewer tokens and steps, not a cut in the headline API rate. Sonnet 5.5 is the second model in the 5.5 family after Opus 5.5. Anthropic says it approaches the stronger model on some structured tasks, allowing buyers to reserve the more expensive Opus for complex judgment and open-ended work.
On Terminal-Bench 4.0, an agentic coding test, Anthropic reports a 70.6% score for Sonnet 5.5 against 10.3% for Sonnet 5 under its published setup. In a test of occupational tasks, it sits two points below Opus 5.5, while the premium model remains stronger on difficult, ambiguous assignments. Anthropic's published API table prices Opus 5.5 at $4 per million input tokens and $20 per million output tokens, double Sonnet's standard prices. The score jump is striking, but task-level economics are more useful for teams paying for agent loops: a model that needs fewer attempts and calls can cost less even when the per-token charge stays constant.
Customers supplied specific, varied early-use results. Slack principal engineer Curtis Allen said its offline Slackbot evaluations improved in fewer steps and with about 14% fewer output tokens, adding that ‘quality and speed are what matter most.’ Zendesk reported support tickets processed 20% faster in its test cases. Balyasny Asset Management said its private suite of 2,441 finance tasks used about 121,000 tokens per answer on the new model versus 497,000 on Sonnet 5, while scoring better. Base44 reported that on 118 app builds Sonnet 5.5 needed 3.6 iterations on average against 7.7 for Opus 5. Those tests are customer accounts under different workloads, not one universal benchmark; together they show where less iteration can alter operating cost.
Anthropic's safeguards reflect a capability increase as well as efficiency. It says the model's cyber ability has advanced enough to use protections like those on Opus 5.5, with higher-risk requests falling back to Sonnet 5. A forthcoming verification route would give approved defenders tiered access to stronger cyber capabilities. The company also added classifiers against industrial-scale extraction of model reasoning and says its roughly 1,850-scenario behavioral audit matched or improved on Sonnet 5 across most safety measures. These controls can affect particular specialist users while leaving routine software development available. The new model's proposition is therefore not just speed; it is a cheaper workhorse with a different set of capability boundaries and routes to approved access.
Sonnet 5.5 is available through Claude products and the API, with Claude Haiku 5.5 expected in coming weeks. The release follows Opus 5.5 by less than a week and lands as Anthropic prepares to discuss its finances with potential public investors. A model family with differentiated prices can steer a large customer to the least costly model that finishes each class of task, rather than forcing a choice between one premium model and a weak budget option. Atlassian said Rovo agents could run up to 30% faster with Sonnet 5.5 than Sonnet 5. For developers, the strongest evidence will be the full cost and completion rate of their own recurring workflows, including retries and tool use.
Analysis
Sonnet 5.5 attacks the cost of a completed task, where fewer tokens and retries can improve both customer return and Anthropic's available capacity. At standard API prices, Opus costs twice as much per input and output token; shifting routine work to a near-premium Sonnet could expand usage without requiring every user to pay for Opus. The Balyasny token comparison is especially large for its private finance suite, although it cannot be applied to all workloads. Anthropic captures value if better efficiency brings more tasks into production and if users keep Opus for genuinely difficult work. The risk is internal substitution: a cheaper model that serves too much of the premium workload can reduce revenue per task unless volume and retention compensate.