Macro breaks micro. Always. In AI, as in crypto, the most consequential forces are invisible to the public order book. This week, SemiAnalysis dropped a speculative grenade: Anthropic has completed training a stronger model—codenamed Mythos 2—but has not released it. Instead, the model is reportedly being used internally to train the next generation, while a separate model called Fable, burdened with extensive safety classifiers, is the public face. If true, this is not a delay. It is a structural shift in how frontier AI capability is accumulated and deployed. The parallel to crypto’s hidden liquidity pools and off-chain settlement is exact: the real value does not flow through the visible API.
The context here is not the usual hype cycle. SemiAnalysis, led by Dylan Patel, is a high-signal source on compute infrastructure. Their claims rest on industry contacts, not official Anthropic confirmation. The model names—Mythos, Mythos 2, Fable—are unverified. But the mechanism they describe is consistent with every major frontier lab’s operational reality. Pre-training and post-training completion do not equal public release. Anthropic’s own ASL (AI Safety Level) framework mandates months of internal red-teaming, capability assessments, and classifier deployment. The gap between "trained" and "shipped" is a standard feature, not a bug. What is novel is the claim that the unreleased model is being used to generate synthetic data—preference pairs, reasoning traces, code verification—to train the next model. This is teacher-student distillation at scale, an engineering-level innovation that decouples capability accumulation from capability exposure.
The core insight is structural: a self-evolution loop forms when an unreleased strong model generates training data for its own successor. This is not a new architecture. It is a pipeline optimization. But its strategic significance is enormous. The strong model’s biases, preferences, and even its unidentified safety flaws propagate through the synthetic data into the next generation. The loop amplifies both strengths and weaknesses, and it does so without any public audit. In crypto terms, this is like a DeFi protocol using its own hidden reserve to backstop a new stablecoin—the reserve is never on-chain, but it determines the peg. The public sees only the output of the loop, not the loop itself. The result is a form of invisible compounding: each generation of model benefits from the previous generation’s internal reasoning, but the quality of that reasoning is never externally validated. The risk is not that the model is unsafe in the traditional sense, but that its safety properties are shaped by a closed feedback cycle that no external researcher can inspect.
The contrarian angle is that the safety narrative is a cover for a deeper competitive strategy. Publicly, Anthropic positions model release delays as a function of responsible AI governance. But the timing aligns with a different logic: holding back the strongest model while using it to supercharge internal products like Claude Code. This is a classic options play. The unreleased model generates no API revenue today, but it creates a moat for downstream products. A developer using Claude Code may not realize they are benefiting from a model that is not available to their competitors. Meanwhile, the safety classifiers on Fable are not just guardrails; they are friction. Increased latency, higher refusal rates, and reduced task completion rates are a tax on the public API. This tax lowers the ceiling of what public users can achieve, widening the gap between what Anthropic can do internally and what the market can access. The hidden implication is that Anthropic’s public API may be deliberately gimped to preserve the value of its internal capability. This is a form of rent extraction that the market cannot price.
The macro consequence is a compression of the competitive window for other labs. If Anthropic’s internal model is already a generation ahead of its public version, then the next public release—when it comes—will represent a discontinuous leap. OpenAI and Google must now compete not just with the current Claude, but with the unknown capability of the next Claude that was trained on data from a model the market has never seen. This recreates the asymmetry of early crypto markets: insiders with access to off-chain liquidity surf the wave while retail only sees the lagging price. The on-chain data—the public API and benchmark results—is a trailing indicator. The real race is in the hidden loop. The impact on enterprise adoption is neutral in the short term, but the long-term dependency risk is high. If a company builds its tools on the public Claude API, it may be locked into a version that is already obsolete relative to what Anthropic itself uses. The switching cost increases with each iteration of the loop.
The takeaway is a call for structural transparency. The crypto industry learned that hidden leverage and off-chain settlement create systemic risk. AI is now facing the same problem. The self-evolution loop is not inherently malicious, but it is inherently opaque. Regulators, investors, and enterprise buyers need to demand not just benchmark scores, but disclosure of internal training pipelines and the use of unreleased models. Otherwise, the market is pricing a product that is a fraction of the actual capability. The next bear market in AI will not be caused by a crypto crash, but by a sudden realization that the public API is a shadow of the internal engine. That realization will force a repricing of every AI stock. The question is not whether the loop exists—it is whether the market will learn to see it before the correction.