Anthropic ships Sonnet 5.5: a faster, cheaper work partner, with its first cyber safeguards

Anthropic released Sonnet 5.5 on September 28, the latest version of its mid-tier model. The company's pitch is a work partner that is significantly faster and cheaper than what came before: specifically, Anthropic claims, about 30% faster than the previous Sonnet 5, with a slower rate of token burn. The lab frames it as an assistant for everyday tasks, coding and office documents included.

The speed pitch

Sonnet 5, 5.5's predecessor, shipped about three months ago with a clear selling point: efficient agentic deployment, meaning the ability to run agents at a lower cost than competitors. With 5.5, the story shifts to speed. Anthropic says the new model is 30% faster than its predecessor and burns tokens significantly more slowly. Put plainly, the same budget should stretch further both in wall-clock time and per completed task.

One caveat before you plan around those numbers: they come from Anthropic. No independent benchmarks for Sonnet 5.5 have appeared alongside the release. For a team running agents at scale, the metric that actually matters is effective cost per completed task, not raw tokens per second, and that number only becomes real once you run your own workloads against the model.

Why the mid-tier model outperforms the flagship

In Anthropic's hierarchy, Sonnet sits below Opus in raw power, but it is built to be more agile. And there is a notable wrinkle in the launch: Anthropic's own benchmarks show Sonnet 5.5 performing better than Opus 5.5 on agentic coding. The company attributes the result to Sonnet's ability to spawn multiple agents without exceeding cost limits, a scenario where a faster, cheaper model can pull ahead over long, multi-agent workflows rather than in single-shot tasks.

Whether that edge holds outside the vendor's test setup is still an open question. If your workloads look like parallel agentic pipelines, it is worth testing; for single-request workloads, the flagship-vs-mid-tier tradeoff may not look the same.

Cyber safeguards for the first time

Anthropic also says Sonnet 5.5 carries what it describes as "comparable" cyber capabilities to Opus 5. The consequence is procedural: 5.5 becomes the first Sonnet model subject to the same cyber safeguards that already apply to Fable and Opus. The company did not elaborate on what those safeguards mean for specific deployments, but the signal is clear: Anthropic now treats its mid-tier model as capable enough to warrant frontier-level restrictions.

A crowded week for model releases

Sonnet 5.5 landed in a busy stretch for the major labs. Just last week, OpenAI released several new models, including enhanced versions of Sol and Luna, its mid-tier and budget-friendly lines. Meta also announced a new model it said will power an upcoming feature tied to its smart glasses. Anthropic added one more line to its own roadmap: a new version of Haiku, its smallest model, is coming in the coming weeks, without a firm date. Read together, the announcements point to where the labs are competing right now: the mid-tier, where most production agent workloads and their budgets actually live.

What to do with it

If you are choosing a model for routine coding or document work, the 5.5 pitch is efficiency: the same class of tasks, more speed, and lower token burn, per Anthropic. The practical next step is unglamorous but decisive: benchmark your own workloads against Sonnet 5 and Opus 5.5 before you switch, and treat every number in this article as the vendor's until an independent test says otherwise.