Cheaper Claude Sonnet 5.5 Launches With Chinese Distillation Defenses
Claude Sonnet 5.5 lands less than a week after Opus 5.5, giving Anthropic a complete, refreshed model lineup right as it courts investors, even as CEO Dario Amodei keeps publicly arguing the industry should slow down.
Anthropic released Claude Sonnet 5.5 on Monday, the second entry in its 5.5 lineup following Opus 5.5’s debut six days earlier, positioning it as a faster, cheaper option for everyday tasks rather than Opus’s complex judgment calls.
The launch comes as Anthropic prepares for a planned IPO, with enterprise clients generating roughly 80% of revenue, per Reuters.
Sonnet 5.5 retains Sonnet 5’s pricing, $2 per million input tokens and $10 per million output tokens, but Anthropic says it uses fewer tokens per task to cut real costs by up to 30%.
A Mid-Tier Model Doing Flagship Work
The company notes, Sonnet 5.5 scored 70.6% on Terminal-Bench 4.0, an agentic coding benchmark, against Sonnet 5’s 10.3%, and scored just two points below Opus 5.5 on GDPval-AA, a test spanning 44 occupations.
Early testers cited by Anthropic described concrete gains: Epic Games said the model handled a system design audit and data flow review at a quality bar it expected from a higher tier.
Balyasny Asset Management reported it used roughly a quarter of the tokens Sonnet 5 needed across 2,441 finance tasks.
That efficiency jump is the real story behind the benchmarks: Sonnet 5.5 batches tool calls more aggressively than its predecessor, meaning fewer steps and lower bills for the exact same output, not just faster raw inference.
Defending Against Theft, Not Just Misuse
Sonnet 5.5 is the first Sonnet model to ship with safety classifiers specifically built to block distillation attacks, where bad actors use thousands of fake accounts to extract a model’s capabilities and build rival systems without Anthropic’s safeguards.
The company also expanded “preserved thinking,” ensuring Claude’s reasoning traces can’t be decoupled from the account that generated them.
Anthropic didn’t name who it’s defending against, but building anti-extraction infrastructure into a mid-tier model, not just its flagship, signals the company now treats reasoning-capability theft as a routine risk rather than an edge case.
The Release Cadence Is the Actual News
Two frontier-family models in six days, with Haiku 5.5 due imminently, is an unusually fast cadence for any AI lab, and a particularly awkward one for Anthropic.
Earlier this month, Amodei urged the AI industry to slow capability releases so safety work could catch up, a stance Anthropic reinforced when launching Opus 5.5 with new containment evaluations.
Sonnet 5.5’s safety section is honest about those limits: Anthropic states plainly that its evaluations cannot reliably catch every failure and that the model may have tendencies its own testing hasn’t found yet.
While positioning a complete three-model lineup before IPO investors is a defensible business strategy, Anthropic is shipping at a pace that would look reckless from any rival, getting away with it simply because they were the ones who first called for a slowdown.
Sources: Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family
Anthropic rolls out second Claude 5.5 model as it builds toward IPO
