This week, Anthropic changed its mind after backlash from developers. The bigger story is what the company still does and why.
The company said it will no longer secretly make Fable 5 less responsive when users ask for help developing Frontier AI models. Instead, Anthropic said these requests will be routed to Opus 4.8, which is a less-good model, and developers will be notified.
This essentially addresses the anomaly to human nature of intentionally giving a bad answer or lying about it. The company said it was “apologetic.”
However, Anthropic still restricts the use of its most powerful public models in certain AI development efforts. The company claims this is about safety and that these restrictions will prevent “foreign adversaries” from using Anthropic’s top-of-the-line models to erode America’s advantages in AI and chips.
That explanation still has its limitations.
These restrictions also help Anthropic protect its business from distillation and information extraction. Rivals then query powerful models, collect their output, and use that data to improve their own systems. These technologies will help open source model providers catch up to Anthropic faster and make Anthropic cheaper in terms of price.
Anthropic warns that Chinese labs are doing this. However, the same threat also comes from open model developers in the US and Europe.
That’s the important point. These restrictions broadly apply to anyone building competing AI models, and Anthropic’s terms of service also prohibit anyone using its products from developing competing products.
Deter open competitors
As such, Anthropic treats Western open model developers the same as Chinese developers. This overturns the idea that it is simply about deterring foreign adversaries. It’s also about deterring competitors.
You can see why. The open model is a strong competitor. An MIT Sloan analysis in January found that open models perform on average 90% better than closed models, and the difference typically closes within 13 weeks. (A year ago, it took 27 weeks to close that gap).
Artificial Analysis tracks the performance of your AI models. This graph shows how well open source models are catching up.
Arena leaderboards show similar pressure. On Thursday, human models still came out on top for specialized text-based tasks like math, coding, and creative writing. But Xiaomi’s no-nonsense MiMo v 2.5 Pro wasn’t far behind.
Graphs showing the performance of various models over time arena
Then look at the price. The cost of the Xiaomi open model is 43 cents per million tokens for inputs and 87 cents per million tokens for outputs. Anthropic’s Fable 5 costs $10 per million input tokens and $50 per million output tokens. According to Arena data, it is at least 20 times more expensive.
When you’re spending billions of dollars building a Frontier model, “about the same” and “much cheaper” must be scary. No wonder Anthropic limits its rivals like this.
“This certainly seems like a business move,” said Nicholas Vincent, a computer science professor at Simon Fraser University who studies how data is used in AI models. “It’s very hard to defend how this could be more safety-focused than business-focused unless you target specific ‘bad actors’ more specifically.”
Anthropic is under no obligation to let competitors take shortcuts in its best technology. But you should be honest about what’s going on. Partly it’s safety, but it’s also business.
Sign up for BI’s Tech Memo newsletter here. Please contact us by email. abarr@businessinsider.com.