Anthropic stops secretly nerfing its top AI model

Imagine paying top dollar for a track car, only to find the manufacturer secretly capped the horsepower every time you tried to study the engine, without warning on the dash.

That is essentially what Anthropic did, and the AI research world is not happy.

The fine print in a 319-page safety card

On Tuesday, the $965B AI powerhouse released Fable 5, its first publicly accessible model from the Mythos class. Anthropic initially teased Mythos back in April, but kept it locked down because it was frighteningly good at bypassing cybersecurity defenses.

Anthropic felt confident enough to launch Fable 5, citing hardened safety guardrails. But buried deep inside a 319-page system card was a landmine: the system would silently downgrade requests related to advanced AI development to a less capable model behind the scenes.

If you were an AI researcher attempting to use Fable 5 to help build or optimize frontier AI, the model quietly pulled its punches without telling you.

Backlash was swift. Fast.ai cofounder Jeremy Howard took to X to sarcastically mock the logic, pointing out that forcing everyone else into degraded tiers under the guise of safety simply guarantees the broader research frontier cannot advance.

Out of the shadows, behind national security

Facing an uprising from technical users, Anthropic folded on the stealth aspect. A company spokesperson admitted they made the wrong tradeoff and apologized for missing the balance.

Starting this week, the stealth rerouting is dead. Flagged requests will now visibly fall back to Opus 4.8. If you are on the API, it will explicitly return a refusal reason every single time it happens.

Still, the restriction itself remains firmly in place. Anthropic defended the practice by pointing to terms of service that ban creating competing models, but also leaned heavily on national security.

According to the company, American and allied leads in frontier silicon and software optimization cannot be compromised. They explicitly want to prevent Claude from being leveraged by foreign adversaries to optimize competing chips, though they insist everyday ML and coding tasks remain untouched.

Timing makes this even messier. Anthropic filed confidentially for an IPO earlier this month. It is also locked in a high-stakes legal fight after Secretary of War Pete Hegseth reaffirmed a defense blacklist against the company over its refusal to support autonomous weapons and mass surveillance.

Why it matters

For builders and researchers, silent downgrades are an absolute nightmare. Debugging complex pipelines is painful enough without a model provider secretly swapping out engine parts under the hood. Forcing Anthropic to make refusals loud and visible is an essential win for API transparency, meaning you can finally catch degraded outputs in your error logs instead of chasing phantom bugs.

For the broader industry, this battle exposes how leading labs are weaponizing safety to protect their moats. Preventing smaller labs from using frontier outputs to bootstrap competitive models has always been standard corporate self-defense, but wrapping that defense in national security rhetoric turns product policy into geopolitical gatekeeping.

We think Anthropic slipped up badly here. Building your brand on safety only works if you practice radical transparency; quietly delivering worse compute while charging full price looks less like ethics and more like deception. While apologizing and exposing the fallback to Opus 4.8 fixes the immediate workflow headache, pulling this stunt right as public markets inspect their IPO credentials leaves a lingering dent in developer trust.