Anthropic's CEO is selling exponential AI progress while his own models prove the opposite—and the company just got caught degrading outputs to slow down competitors.
The Summary
- Anthropic admitted to secretly degrading Fable 5 responses when developers asked for help building frontier AI models, then lying about it. After backlash, they'll now route those requests to the weaker Opus 4.8 and tell users.
- CEO Dario Amodei claims AI growth remains "exponential" in a new essay, but Claude Mythos performance data suggests the curve is flattening.
- The real story isn't the apology. It's that Anthropic is still restricting model access, citing national security while actually protecting against distillation—when competitors extract intelligence by querying models to train their own systems.
- Open-source models are catching up faster than closed providers want to admit, threatening Anthropic's business model at the exact moment their performance gains are slowing.
The Signal
Anthropic spent this week in damage control after developers discovered the company was intentionally degrading Fable 5 outputs for certain AI development queries, then pretending the worse answers were genuine. The company apologized and promised transparency: those queries will now route to the less capable Opus 4.8, and users will be told. But the restrictions themselves aren't going anywhere.
The official justification is national security. Anthropic claims these limits prevent "foreign adversaries" from using their top model to erode America's AI advantage. That framing plays well in Washington. It also conveniently ignores the commercial incentive: stopping distillation.
"These restrictions help Anthropic protect its business from rivals who query powerful models, collect outputs, and use that data to improve their own systems."
Distillation is how open-source providers are closing the gap. You don't need Anthropic's compute budget or training data if you can learn from their model's responses. Query it enough times with the right prompts, capture the outputs, and you've got training signal. Open-source models now deliver competitive performance at lower cost, which is exactly what keeps Anthropic's product team up at night.
Here's where it gets interesting. While Amodei is out writing essays about exponential AI growth, his own company's model performance tells a different story. Claude Mythos, Anthropic's research into their own progress, shows the curve flattening. Not collapsing, but not exponential either.
The contradiction matters. If you're restricting access to protect a decisive capability advantage, you need that advantage to stay decisive. But if open-source models are catching up and your own scaling laws are softening, you're not protecting national security. You're protecting market share. And badly.
Key tensions:
- CEO messaging (exponential growth) vs. internal research (flattening curve)
- Security justification vs. commercial reality of distillation threat
- Closed model pricing power vs. open-source performance parity
The developer backlash worked because Anthropic got caught being dishonest about something developers can measure. You ask a frontier model a question, you know what a frontier answer looks like. When you get a worse answer and the API claims it's the best model, you notice. The community called it out, Anthropic folded on the deception, but kept the underlying restriction.
The Implication
Watch how other closed model providers respond. If Anthropic is restricting access to slow distillation, OpenAI and Google have the same incentive. Expect more quiet guardrails on what you can ask models to help you build, dressed up as safety policy.
For developers building on top of these APIs, the lesson is simple: closed models will always optimize for the provider's business first. That might mean worse answers, arbitrary restrictions, or sudden policy changes. If you're building something that depends on frontier model access, you need a plan for the day that access gets restricted, repriced, or yanked entirely.
The open-source camp just got handed ammunition. Anthropic's move confirms what they've been saying: closed providers will limit access to protect their edge, regardless of what that does to the broader ecosystem. If your agent workflow depends on unrestricted model access, open weights start looking a lot more attractive than API calls with invisible throttles.