Anthropic Released Another Claude Model — Days After Its Own CEO Called for an AI Slowdown
On September 28, Anthropic shipped Claude Sonnet 5.5. It's the second model launch from the company since CEO Dario Amodei published an essay just 16 days earlier arguing "we must slow the pace at which we improve the capabilities of AI models."
The apparent contradiction is worth unpacking, because Anthropic's own explanation reveals something more interesting than a simple say-one-thing-do-another story:
→ Sonnet 5.5 is faster (30%+) and cheaper (up to 30% less) than its predecessor, built for everyday business and coding work — not a capability frontier push
→ Anthropic explicitly stated the model "doesn't advance the frontier" of what its models can do, and scoped most of its alignment testing to a narrower set of risks that apply "to models of any capability level"
→ It scores 70.6% on Terminal-Bench 4.0 and matches Opus 5.5 (Anthropic's actual frontier model) on knowledge benchmarks, often using fewer tokens to get there
→ Available immediately across Claude.ai, the API, AWS, Google Cloud, and Azure
Here's the part that actually complicates the slowdown debate: even a "routine," non-frontier model release is now reaching cyber-relevant skill levels fast enough to matter. Cybersecurity stocks rallied the same week Amodei's essay sank AI stocks broadly — firms like CrowdStrike and Palo Alto are already brought in to test unreleased Anthropic and OpenAI models before launch. That verified-access model works, but it also concentrates pre-release scrutiny in the hands of the biggest security vendors, raising its own question about who gets to see dangerous capabilities first.
My read: Amodei's "slow down" was never a call to stop shipping — it was a call to stop racing on raw capability while safety testing lags behind. Sonnet 5.5 looks like Anthropic trying to prove those two things can be decoupled: keep releasing useful, cheaper models for real work, while reserving the "frontier" label — and the heavier safety bar that comes with it — for models that actually earn it. Whether the rest of the industry buys that distinction, or just sees a slowdown pledge with an asterisk, is the real test.
Is "we're pacing the frontier, not pausing shipping" a meaningful distinction to you — or just a convenient one?
#Anthropic #ClaudeAI #AISafety #ArtificialIntelligence #TechNews #DarioAmodei
— 𝔖𝔞𝔫𝔡𝔢𝔢𝔭 ℜ𝔞𝔦𝔷𝔞