Anthropic uses AI to speed up its own model development

Anthropic has begun using AI systems to automate parts of its own AI development process. The safety research group METR reports that Anthropic’s models are now improving at about twice the speed they did last year.

Why it matters

The METR report describes Anthropic’s shift from using AI for routine work to letting AI systems automate the process of building new, more capable models. Over 11 weeks, METR found that this feedback loop is making model improvement much faster. Researchers quoted by The Verge say this could make oversight harder and risks less predictable, since automation can amplify mistakes or unexpected behaviors.

For businesses that rely on AI, this means the development cycle for new models could become much shorter. If other companies adopt Anthropic’s approach, the time between major model releases may shrink, making it harder for competitors and regulators to keep up. These concerns are similar to those seen in recent market reactions covered in our report on AI industry fears and stock selloffs.

Anthropic’s automated development process under review

METR’s 11-week investigation focused on how Anthropic uses AI tools to automate model training and evaluation. The report says Anthropic’s automation has roughly doubled the speed of model improvement. METR does not name specific models or codebases, but describes a feedback loop where AI-written code and AI-run evaluations help produce better-performing models, with humans still involved but less central than before.

The Verge quotes researchers who are concerned that this kind of automation could make it harder to audit or understand how models are trained and improved. There is also a risk that errors or unexpected behaviors in one generation could be amplified in the next. METR’s report suggests Anthropic is aware of these risks but continues to automate because the productivity gains are significant. The speed of adoption has surprised some in the AI safety community, who expected a slower shift.

What this means if you run a business

For companies that depend on AI models for products or operations, Anthropic’s move signals that model capabilities may change faster and with less warning. A model that works reliably today could be replaced or behave differently sooner than expected. This increases the pressure on technology partners, procurement teams, and compliance processes to keep up with changes.

If your business relies on commercial AI APIs, you may see models updated or swapped out more quickly, sometimes with less documentation or notice. This means review cycles and risk assessments may need to be more frequent. For companies developing their own AI, Anthropic’s approach could prompt questions about automating parts of your own pipeline, but also about how to maintain oversight and safety checks.

Anthropic’s rapid automation may also push competitors to follow suit, making the AI market less predictable. For more analysis on the business risks tied to accelerated AI development, see our report on AI stock selloff and data center costs.