SemiAnalysis dropped a single thread. It claimed Anthropic has a stronger model—Mythos 2—completed but not released. The industry lit up. Not because of the model. Because of the mechanism.
The rumor isn't the story. The story is the second layer: that Anthropic is using this unreleased model to train the next generation internally. This is a closed loop. The public never sees the strongest model. But the strongest model shapes the next one. Code does not lie, but incentives do.
Let me be clear. I have no inside access to Anthropic's labs. I don't know if Mythos 2 exists. But I know the architecture of this dynamic. And it's not new. It's a pattern I've seen in crypto audits for years. The building of a system where the public version is a facade, and the real value is hidden behind a wall of safety and access controls.
Context: The Safety Theater
Anthropic's AI Safety Level (ASL) framework is the stated reason for delays. Model training completes. Then comes months of internal evaluation, red-teaming, and classification layer deployment. The narrative is responsibility. The reality is a bottleneck.
But here's what the rumor suggests: that bottleneck isn't just a gate. It's a filter. The strongest model stays inside. The weaker model—or a neutered version—goes to the public. And the strongest model is used to generate synthetic data for the next version. This is a teacher-student distillation loop. The public gets the student. The company keeps the teacher.
This is technically feasible. Industry precedent is clear. GPT-4 generated SFT data for Alpaca. DeepSeek-R1 generated reasoning traces for smaller models. The math works. The question is not can it be done. It is: what is the cost of the veil?
Core: The Structural Deconstruction
Let me walk through the mechanics. A model like Mythos 2, if it exists, is trained on massive compute. That compute has a cost. If the model is not released, that cost cannot be recovered through API revenue or subscription fees. The ROI window shifts. The books take a hit. But the model is not idle. It is used to generate high-quality preference data, reasoning traces, and code verification samples. This data is then used to train the next model.
This is a form of capital lock-up. The company invests in a capability that is not immediately monetized. But it becomes a competitive lever for the next generation. The strongest model is not sold. It is used to make the next strongest model. And the public never sees the gap.
From my experience auditing the 0x Protocol v2 vulnerability in 2017, I learned that the most dangerous flaws are not in the code that is visible. They are in the assumptions about what is hidden. The same applies here. The public assumes that the released model is the best. The rumor suggests otherwise. The vulnerability is not in the model. It is in the trust.
I read the reverts before the headlines. In this case, the revert is the public's perception of frontier capability. The supposed 'safety' delay creates an information asymmetry. The company knows its best. The market does not. This is a centralization of knowledge that mirrors the oracle feed latency problem in DeFi. The truth is there, but the feed is delayed.
Trace the gas, find the truth. Here, the gas is the compute. The internal training loop consumes compute. That compute is a cost that does not show up in public earnings reports as a line item for 'Model Mythos 2.' It is buried in R&D. The market cannot price it. The auditor cannot verify it. The logic held until the liquidity dried up. Here, the liquidity is transparency.
Contrarian: What the Bulls Got Right
I want to pause. Not everything about this dynamic is malicious. There is a legitimate argument for keeping a model internal. The safety classification layers required for a powerful model are significant. Deploying it publicly could lead to abuse. The company has a duty to mitigate risk. The delay is a trade-off, not a failure.
Moreover, using an internal model to train the next generation is a strategic asset. It allows the company to build a moat. The competition cannot see the intermediate steps. The end result is a sudden jump in capability. This is not a Ponzi. It is a legitimate R&D strategy.
But the problem is the asymmetry. The public is funding the compute through API costs. The investors are funding the compute through equity. And the strongest model is hidden. The 'safety' narrative becomes a cover for a competitive advantage. This is not a conspiracy. It is an incentive misalignment. The company's incentives are to maximize its own power. The public's incentives are to understand the true frontier. These are not aligned.
Silence is just uncompiled potential energy. The potential energy here is the knowledge gap. The model exists. The data is generated. The next model is trained. And the public is left with a Fable—a mythos—while the real strength is locked away.
Takeaway: The Accountability Call
This is not an attack on Anthropic. It is a call for transparency. If the rumor is true, the industry needs to understand the implications. The strongest model should not be a secret. The safety classification should be public. The internal training loop should be disclosed. The market needs to price this correctly.
Entropy always wins if you stop watching. The industry is watching the wrong thing. It is watching the public model. It should be watching the hidden one. The truth is not in the headlines. It is in the bytes. Read the bytes.