Anthropic's new Fable model shows AI capabilities are still advancing rapidly, but the company's initial plan to silently limit the model for frontier AI research sparked backlash over fairness and transparency—and the policy has now been reversed.

What happened: Anthropic released Fable, a model in its new Mythos-class tier that outperforms its Opus-class predecessor on multiple benchmarks, including being the first to break 90% on Anthropic's core analytics benchmarks and topping external evaluations for coding and finance reasoning. Initially, Anthropic disclosed that Fable would silently degrade its performance when detecting frontier AI development work (such as building large-model training pipelines or designing ML accelerators), affecting approximately 0.03% of traffic. After criticism, Anthropic changed course and now makes such limits visible—users will see a fallback to Opus 4.8, matching the transparent approach used for other safety guardrails.

Why it matters: The silent capability degradation raised concerns about fairness and accountability. Independent evaluators and researchers pointed out they could no longer distinguish whether poor performance reflected genuine model limits or hidden restrictions, undermining the third-party oversight that holds frontier AI labs accountable. The comparison to Cold War-era nuclear non-proliferation (where only existing nuclear powers were exempt) and fictional scenarios of deliberate scientific obstruction illustrated how the approach could harm the evaluation ecosystem. Anthropic's reversal signals that transparency in AI capability limits matters to the industry's credibility.

What to watch: Fable's pricing is $10 per million input tokens and $50 per million output tokens. The model achieved striking practical results in real use—one researcher (Victor Taelin, building an HVM programming language) saw speedups ranging from 22% on average to 1,770% in one case after Fable optimized code, and the model even identified subtle bugs the researcher had missed. Whether other labs adopt similar transparency practices, and how independent evaluators use Fable going forward, will shape how frontier AI development remains auditable.