Anthropic launches Sonnet 5.5: 30 percent faster

with cyber capabilities that push it to Opus-level safeguards Anthropic released Sonnet 5.5 yesterday, the latest version of its mid-tier model, just three months after Sonnet 5 debuted. The company positions the model as a significantly cheaper and faster work partner for everyday tasks, code writing, document generation and running agents. The headline claim: a 30 percent speed improvement over the previous generation, alongside a much lower token burn rate, which should allow agents to run longer without blowing the budget.

In Anthropic's hierarchy, Sonnet sits below Opus, the flagship model, but moves faster. According to the lab's internal benchmarks, Sonnet 5.5 now outperforms Opus 5.5 on agentic coding. The explanation: the ability to run multiple agents in parallel without breaching cost limits. That claim deserves caution; the tests are the manufacturer's own, on scenarios the company selected, and no independent evaluation has yet confirmed the gap holds in practice.

The more interesting development is Anthropic's statement that Sonnet 5.5 has "comparable cyber capabilities" to Opus 5. As a result, this is the first Sonnet model subject to the same safeguards applied to Fable and Opus. In practice that means stricter usage limits, tighter monitoring, and likely more frequent refusals on requests related to vulnerability exploitation or offensive code writing. Anthropic did not specify which cyber tests were performed, and there is no public definition of "comparable," so it is difficult to judge whether this amounts to full parity or only partial overlap.

In the background, Anthropic confirms a new version of Haiku, the family's small, low-cost model, will arrive "in the coming weeks" without a firm date. The market is not waiting: last week OpenAI launched improved versions of Sol and Luna, its mid-tier and budget counterparts, and Meta announced a new model that will power a future feature in its smart glasses. The release cadence across every major lab over the past year makes any "significant announcement" highly provisional, valid only until the next wave arrives.