Anthropic Mythos: 4 reasons the company is not releasing its strongest model

Can a single AI model be so powerful that the company that built it is afraid to release it publicly? It can.

Anthropic disclosed on April 7 the existence of a model it calls Mythos internally. It is more capable than Opus 4.7, but it is not on sale to ordinary users. It is available only through Project Glasswing, a program limited to around fifty cyber defense partners, including some US government and allied agencies.

This is a rare case of an AI company publicly admitting it is holding back a stronger model than the one it sells. The stated reason is the risk of misuse for attacks on critical infrastructure. A year ago, almost every lab dismissed that scenario as unlikely.

1. What Mythos does better than Opus 4.7

Mythos is significantly stronger in mathematics and especially in cybersecurity. Concretely:

• Finding vulnerabilities in code (flaws that could be exploited in an attack)
• Describing an attack vector (how a specific vulnerability could be used)
• Proposing defenses

On the defensive side, that is an excellent tool. The problem is that the same capability can be used to find vulnerabilities as part of an attack.

Anthropic argues that Mythos audits secure code more effectively than Opus 4.7. The company cannot guarantee that the same model in the hands of a state actor would not push in the opposite direction.

2. Project Glasswing as a controlled test

Glasswing is the program through which Mythos goes out under limited partnerships. According to publicly available information, participants are around fifty organizations specialized in cyber defense.

Anthropic plans to widen the circle to more US government agencies and allied countries before considering any public release of the model.

The full program terms have not been published. What is known:

• Partners go through vetting
• They sign usage agreements
• They cannot lend access to third parties

3. Users accidentally spotted traces in standard tools

In late May, users of Claude Code and Claude Security tools briefly saw an option to enable the Mythos model. The toggle was pulled immediately.

This suggests Anthropic is wiring infrastructure for a wider rollout. But the safety mitigations are not yet at a level the company considers acceptable.

Anthropic has publicly admitted that no organization, not just them, has developed safeguards strong enough to prevent misuse of a model of this caliber. The standard itself still has to be defined.

4. What this means for the AI market

The split between a “frontier” research model and a “production” model for the public is becoming the new standard:

• Google with certain Gemini variants
• OpenAI with internal GPT versions
• Now Anthropic with Mythos

The market is settling into a pattern where the strongest versions are not for sale.

For users outside the industry, this means a 20-dollar monthly subscription is not necessarily getting you the strongest model. The gap is widening. That said, Opus 4.7, which is available to everyone, already covers almost everything a normal user needs.

What this means for users in Serbia

Mythos is not available to users in Serbia, neither through claude.ai nor through the API.

Opus 4.7 and Sonnet 4.6 are the only options, and both work without geographic restrictions on Serbian payment cards.

Development teams that would like to apply for Project Glasswing have to be specialized cybersecurity firms with established references. Anthropic is not making exceptions for individual developers.

Is it a good idea for a company to keep its strongest model to itself?

The argument for: misuse control and gradual learning before wide release.

The argument against: a monopoly on a tool that could move entire research disciplines.

Neither approach is safe. Caution is probably better than haste, especially when the model in question could meaningfully help attacks on infrastructure.

Would you pay a higher subscription to get access to a company’s internally strongest model?