What Exactly Did They Cap?

The two restricted areas aren’t obscure esoterica. They’re the two most critical skills for building better AI systems:

1. Accelerator kernel optimization, the low-level code that makes GPUs and other chips run training and inference efficiently
2. Improving large language model architectures, literally, the ability to design better transformer structures

In both domains, Fable 5.2 and Mythos 5.2 perform at roughly the level of Claude Opus 5.5. That’s not a coincidence. That’s a ceiling Anthropic deliberately installed.

The company says it “ensured that the training changes did not affect related capabilities like agentic coding, hardware development, and traditional AI or machine learning research (including interpretability research).” So the model can still be exceptional at writing applications, designing robots, and doing physics. It just can’t meaningfully contribute to improving the models that will replace it.

As reported by MindStudio, this is a targeted restriction rather than a blanket one. A model can be exceptional at coordinating agents on complex software projects and still be deliberately mediocre at the one thing that would let it help build its own successor.

The stated motivation is recursive self-improvement (RSI), the scenario where an AI system becomes capable enough at AI research to meaningfully contribute to building its own successor. Once that loop closes, progress stops being bounded by human headcount and starts being bounded only by compute and model capability. That’s a fundamentally different scaling regime, and it’s one Anthropic apparently isn’t ready to let anyone outside their walls access.

The “Safety” Argument vs. Reality

Here’s where things get spicy. The official line is safety and alignment. The community’s response has been… let’s call it skeptical.

The prevailing sentiment on developer forums is that this has less to do with existential risk and more to do with competitive advantage. Kernel optimization is precisely what lets developers run smaller models efficiently on their own hardware, or what lets a competitor squeeze more performance out of their chip clusters without buying more GPUs. Capping that capability isn’t just about preventing RSI, it’s about making sure the people building the next frontier model need Anthropic’s infrastructure to do it.

One commenter captured the frustration perfectly: kernel optimization “helps us mere plebs run small local models on old hardware outside their sandpit but not relevant to ‘dangerous’ frontier progress.” In what universe is optimizing code for a 10-year-old GPU a threat to humanity?

The answer, as several developers pointed out, is that it’s a threat to the bottom line. If Fable 5.2 could optimize CUDA kernels effectively, it would make open-weight models more competitive. It would let smaller labs train and run models more cheaply. It would flatten the infrastructure advantage that frontier labs currently enjoy.

Wait, They Capped Kernel Optimization? Really?

Yes. And if you work with CUDA, HIP, or any accelerator-level code, this affects you directly, even if you’ll never touch Fable 5.2 yourself.

The cap extends to “kernel development on certain nonstandard accelerators”, with Anthropic implementing “transparent fallbacks to Opus 5” for those topics. So if you ask Fable 5.2 to help you write a custom kernel for a Chinese accelerator or an older GPU architecture, you don’t get Fable-level help. You get whatever Opus 5 can do, routed through a model that’s theoretically much more capable but isn’t allowed to show it.

That’s a massive usability difference for anyone working on inference optimization, edge deployment, or non-NVIDIA hardware. And it’s one of the reasons that Fable 5’s underwhelming real-world performance despite benchmarks has been such a persistent community complaint, the gap between what the model demonstrates internally and what it does for external users is growing.

Several developers have also reported that Anthropic’s models trigger safety guardrails when they see chain-of-thought from other models, making it a “massive pain” to do interpretability research using Claude. If you’re trying to study how other AI systems reason, the tool you’d most want to use actively blocks you from seeing the reasoning.

The Internal vs. Public Model Gap

The capping wouldn’t be so galling if Anthropic’s internal models were equally restricted. But the evidence, from the METR report and subsequent analysis, points to a different reality.

The report describes internal checkpoints labeled “model 2”, “model 3”, and “model 4.” Fable 5.2 and Mythos 5.2 reportedly correspond to “model 3 early.” That means the public is using a version that sits behind Anthropic’s internal frontier, and the gap, while described as “razor thin” in overall capability, is presumably wider in the specific domains that were deliberately capped.

The first internal model to exceed 50% on Anthropic’s AECI (AI-Enabled Capability Improvement) benchmark was model 4. That model isn’t public. Fable 5.2, with its deliberately limited training, runs behind it.

The not-so-subtle implication is that Anthropic’s internal systems are already operating beyond the RSI threshold they’re trying to prevent the public from approaching. The company gets the research acceleration. The rest of us get a model that’s exceptional at everything except the things that would let us compete.

This is the fundamental tension at the heart of the debate over monolithic vs. modular AI design: the more concentrated AI capability becomes, the more decisions about what’s “safe” to release become decisions about who gets to have what power.

The Distillation Angle: Protecting the Moat

The capability capping isn’t the only way Anthropic is trying to control the spread of AI capability. The company also announced its “strongest-ever protections against distillation attacks”, the practice of training a new model on a stronger model’s outputs, including intermediate reasoning, to extract its capabilities.

Anthropic reports that distillation “occurs illicitly on an industrial scale, often involving thousands of fake Claude accounts.” They’ve deployed techniques that they claim have “reduced the data harvested by distillers by 90% to 99%.”

There’s also the “preserved thinking” API feature, which prevents distillers from editing Claude’s prior context to extract its reasoning. And there are persistent complaints about reasoning traces from Anthropic models appearing disjointed or referencing tools that don’t make sense, behavior that’s consistent with intentional obfuscation rather than model incompetence.

If the reasoning traces are deliberately scrambled to make them harder to harvest and imitate, that’s a direct attack on the open-source ecosystem’s ability to learn from frontier models. Open-weight models that perform comparably on leaderboards often do so partly because of distillation from closed models. Cut that pipeline, and the pace of open-weight model innovation versus closed frontier models could slow dramatically.

What Does This Actually Mean for Users?

For most practical purposes, Anthropic says, the capping is invisible. Fable 5.2 handles coding, agentic workflows, scientific research, hardware design, and robotics without any sign of being held back. The model’s benchmark numbers confirm this: Fable 5.2 leads or ties nearly every major benchmark, including significant jumps in agentic scientific research (74.6% vs. 58.7% for Opus 5.5) and robotics (64.9% vs. 39.4%).

The restriction only becomes visible when you push the model toward the specific domains of kernel optimization and architecture improvement. There are reported speed-run benchmarks, like nano GPT training, where Fable 5.2 and Opus-class models pull ahead of competitors. But the model’s performance on FrontierCode 2.0 shows an interesting pattern: it peaks at “extra high” reasoning effort and then declines at the maximum compute setting, at a noticeably higher cost. That’s an unusual profile that hints at engineered limits rather than natural model behavior.

There’s also a cost question worth flagging. Artificial Analysis measured an average task cost of roughly $26.54 at max reasoning settings, making Fable 5.2 the most expensive model in its comparison set at full power. API pricing runs $8/$40 per million input/output tokens, cheaper than Fable 5.1’s $10/$50 but double Opus 5.5’s per-token rate. And Anthropic itself recommends Opus 5.5 as the daily driver for most tasks, positioning Fable as an advisor or orchestrator for the hardest problems.

So the model that was deliberately held back in critical capabilities is also the one you’ll pay the most to run. The people who genuinely need frontier AI research acceleration, the ones who would benefit most from uncapped kernel optimization and architecture skills, are precisely the ones who can’t access them without paying a premium for a model that’s still missing exactly what they need.

The Competitive Landscape

The open question is whether this strategy holds up. Other labs haven’t publicly adopted similar capability caps, and there’s a real risk that Anthropic just handed its competitors a roadmap.

If Google or OpenAI releases models without these training-time restrictions, their models could pull ahead specifically in AI-research-relevant skills. That would be immediately visible in specialized benchmarks, and it would put Anthropic in the awkward position of explaining why its “safety first” approach made its models less capable than the alternatives.

The company hinted at this concern in its own guidance, noting that competitors’ models “could in principle pull ahead specifically in AI-research-relevant skills” if they don’t apply similar restrictions.

Meanwhile, the broader industry is moving in the opposite direction. OpenAI’s recent shift to specialized dual-model architecture suggests that capability differentiation is becoming the norm, but it’s differentiation toward more specialized capability, not deliberate restriction. And smaller, more efficient models continue to make the compute-efficiency case that Anthropic’s capping ironically undermines.

The Deeper Question: Who Gets to Decide?

Setting aside the competitive dynamics, the capability capping raises a more fundamental question that the AI community hasn’t adequately wrestled with: who decides what capabilities are too dangerous to train?

Anthropic made this decision unilaterally, without public consultation, without regulatory oversight, and without a transparent framework for what gets capped and why. The company positions it as a safety measure, but the line between “safety” and “competitive moat” is blurrier than they’d like to admit.

The broader pattern of humans deliberately restricting AI capabilities, whether for safety, to protect jobs and professional authority, or for competitive advantage, is likely to become a defining theme of the next several years. The EU AI Act requires human oversight on high-risk systems. Self-driving cars are legally required to have a human ready to intervene. Content moderation systems are deliberately throttled to avoid overconfident answers.

But there’s a difference between regulatory guardrails that apply equally to everyone and a private company deciding that its models won’t be allowed to improve AI systems, while its internal models presumably can. The former is democratic governance. The latter is the flattening of knowledge hierarchies being decided by whoever controls the most powerful systems.

The Bottom Line

Fable 5.2 is extraordinary. It built an operating system in eight days. It formalized Fermat’s Last Theorem in 30 hours. It’s the most capable public model ever released, and it’s not particularly close.

But it’s also a reminder that the frontier of AI capability isn’t just a technical question, it’s a political one. The people who build the most powerful systems are also the ones deciding what the rest of us get to access. They’re drawing lines around kernel optimization and architecture research, framed as safety measures, implemented as competitive moats, and enforced through training-time restrictions that can’t be undone by prompts, fine-tuning, or inference-time prompting.

The genuinely unsettling thing isn’t that Anthropic capped these capabilities. It’s that they’re right to be worried about what uncapped models could do, and the rest of us have no say in how that worry gets translated into policy.

For developers and researchers, the practical takeaway is straightforward: if your work depends on pushing the absolute frontier of AI research capability, you need to plan for the possibility that the most capable public models will always be deliberately less capable than they could be. Your alternatives are open-weight models that trade raw capability for transparency, or frontier models that come with invisible ceilings.

The theoretical questions, about recursive self-improvement, about concentration of power, about who gets to decide what “safe” means, don’t have easy answers. But they’re now concrete enough that pretending they’re hypothetical is no longer a serious option.