Anthropic launches Claude Sonnet 5.5, faster and cheaper than the previous version

Anthropic launched Claude Sonnet 5.5 this Monday (28th), a new version of its mid-tier model, promising gains in speed and cost compared to its predecessor. According to TechCrunch, the company describes Sonnet 5.5 as an ideal assistant for everyday tasks, including code generation and office document production.

The previous Sonnet 5 had been announced about three months earlier, focused on efficient agentic deployment: the promise was to run agents at a lower cost than competitors. With 5.5, the pitch shifts axis. Now the selling point is speed: Anthropic states that the new model is 30% faster than its predecessor and burns tokens at a significantly lower rate.

Sonnet More Agile Than Opus in Agentic Coding

In Anthropic's model hierarchy, Sonnet occupies the mid-tier layer, below Opus in raw capability. But the company claims that, in certain scenarios, it is more useful precisely because it is more agile.

The benchmarks released by Anthropic show Sonnet 5.5 outperforming Opus 5.5 in agentic coding tasks. The explanation pointed out by the report is Sonnet's ability to fire off multiple agents in parallel without blowing past cost limits, something that becomes harder to sustain with a model that is more expensive per token.

In short: for those building code agent pipelines, this flips a common logic. It isn't necessary to move up to the more expensive model to get the best result on a specific task; sometimes the cheaper model, run at greater volume and parallelism, delivers more.

Cybersecurity Capabilities Under New Safeguard

Anthropic also states that Sonnet 5.5 has cyber capabilities "comparable" to those of Opus 5. Because of this, 5.5 becomes the first model in the Sonnet line subject to the same cybersecurity safeguards that already applied to the Fable and Opus models.

The source does not detail the technical content of these safeguards, but the fact that a mid-tier model crosses this threshold is relevant for anyone integrating Claude into offensive security workflows, automated pentesting, or vulnerability analysis. Teams that use Sonnet for this type of task may start encountering restrictions or additional controls that previously only appeared in the top-tier models.

What's Left Behind and What Comes Next

Anthropic has not yet given a release date for the next version of Haiku, its leanest and cheapest model, but confirmed it is expected "in the coming weeks". This completes the line's update cycle: Opus, Sonnet, and soon Haiku, all refreshed within a short span of months.

The move does not happen in isolation. In the week before Sonnet 5.5's launch, OpenAI updated Sol and Luna, its mid-tier and low-cost models. Meta, for its part, announced a new model intended to power a future feature of the company's smart glasses. The pace of releases from major AI labs has remained steady over the past year, with each one repositioning its mid-tier layer as the most fiercely contested battleground.

What Changes for Those Building with Claude in Brazil

For teams already using the Claude API in production, Sonnet 5.5's cost and latency improvements have a direct effect on the bill. A model that spends fewer tokens per task and responds faster reduces the operational cost of running agents continuously, something that weighs doubly for Brazilian teams paying their invoice in dollars.

The data on agentic coding also deserves practical attention. If Sonnet 5.5 outperforms Opus 5.5 in agentic code tasks, it makes sense to reassess which model is set as the default in tools like Claude Code or in API integrations that currently point to Opus by default, assuming that "more expensive" means "better result".

The tightening of cybersecurity safeguards, meanwhile, is a point of attention for those using Sonnet in offensive security contexts or penetration-test automation. It's worth following Anthropic's official documentation to understand whether this implies new usage limits, additional verification requirements, or changes to the terms of service for this type of application.

What remains open is the detailed per-token pricing and the exact latency numbers in independent benchmarks, since TechCrunch's report only reproduces Anthropic's own claims. Only the practice of those already running the model in production will confirm whether the speed and cost gains hold up beyond the figures released by the company.