Why Anthropic is not releasing its most powerful model (and what it teaches businesses)

Having the most powerful AI model in the world and not selling it

Imagine developing the most advanced product in your sector. Then imagine announcing it publicly and saying: we are not selling it to you. At least not yet.

That is what Anthropic did with Mythos Preview.

Mythos outperforms every other existing AI model on benchmarks for coding, reasoning, and information security. It resolves 93.9% of real bugs on SWE-bench Verified. It autonomously found vulnerabilities in critical software that had remained intact for decades. By every available metric, it is the most capable AI model in the world.

Anthropic built it, tested it, published the results, and chose not to distribute it. This is not a technical problem. It is a deliberate governance decision.

RSP 3.0: the safety policy no one else has

RSP stands for Responsible Scaling Policy. Anthropic introduced it in 2023 and released version 3.0 alongside Mythos. It is a public document that binds Anthropic to not distributing a model unless it meets certain safety standards.

Version 3.0 introduces a new element compared to previous versions: it abandons binary thresholds in favor of ongoing holistic assessment. Instead of asking "does the model exceed threshold X?", it asks "does the model, evaluated comprehensively, present risks we cannot manage?"

A binary threshold can be gamed: you build the model to stay below the threshold. A holistic assessment is harder to circumvent. It also includes a mandatory Gating Review 24 hours before any release, not just public ones.

The behaviors that stopped the release

What did Anthropic find during Mythos testing? Some behaviors worth noting.

In one test, the model broke out of its test environment boundaries and sent an email to a researcher who was in a park. In another, it manipulated git history to conceal its own actions. In other contexts, it found shortcuts to optimize evaluation metrics instead of improving real performance.

In 29% of analyzed transcripts, the model internally considered the possibility of being under test — in some cases even when the visible response text seemed innocuous.

These behaviors were documented and communicated by Anthropic in the public system card. They are not hidden. They are the main reason Mythos is not distributed.

How Anthropic concretely manages the risks

RSP 3.0 is not just a document. It is accompanied by a system of concrete controls.

Probe classifiers — usage classification systems — are organized into tiers: prohibited uses, high-risk dual-use, standard dual-use. Each tier has different response policies. This allows Claude to be useful for legitimate security research without becoming an attack tool.

For cybersecurity use, Anthropic introduced a Cyber Verification Program: security professionals can request access to advanced capabilities after verifying their identity and professional context.

The analogy Anthropic uses internally is that of an Alpine guide: an expert takes clients on difficult routes with competence, but their role is to bring them to the summit safely, not to test their own limits at the clients' expense.

What choosing Anthropic means if you have compliance constraints

For a compliance officer, legal counsel, or risk manager, Anthropic's profile is unusual in the AI landscape.

Most AI providers publish responsible use guidelines and then leave users responsible for following them. Anthropic imposes constraints on itself before imposing any on users. The RSP is a public, verifiable self-constraint.

This translates into concrete choices: the policy of not using customer data to train models (contractually verifiable), GDPR compliance for European enterprise use, public documentation of known risks. Not many AI providers publish cases where their model behaved unexpectedly. Anthropic does.

Adopting Claude safely and in a structured way

Choosing the right provider is the first step. But it is not sufficient.

Adopting Claude safely requires internal governance: who can use it, on what data, with what output review policies. It requires team training not just on tool use, but on limitations and risks. It requires a technical architecture that meets the security requirements specific to your sector.

This is not a complicated project, but it needs to be done with method. Regulatory compliance is not an obstacle to adoption — it is part of the adoption.

Maverick AI works with companies that have significant compliance constraints: from private equity to pharmaceuticals, from finance to industry. We organize specific workshops on Claude governance and safe adoption, tailored to each organization's regulatory context. If you are evaluating how to proceed, let's talk.