Anthropic Holds Back More Powerful AI Model Despite Continued Development

Anthropic has developed an internal AI system more capable than its flagship Mythos model but has decided not to release it publicly, according to the company's latest risk assessment report issued Friday.

The unreleased system, designated "Model 2," demonstrates "noticeable improvement for internal tasks" and is being used extensively within Anthropic alongside Mythos 5 for coding, autonomous agent work, and data generation. However, the performance gains don't match the substantial leap seen earlier this year when Anthropic moved from Opus 4.6 to Mythos.

Why it matters

As competitors like OpenAI delay releases due to safety concerns, Anthropic's decision to continue aggressive internal development while withholding public deployment could position it to reach artificial general intelligence first—a scenario that raises questions about competitive dynamics in frontier AI development.

Rising risk assessments

Anthroptic elevated its overall risk estimate for AI misalignment in high-stakes scenarios from "very low" to "low," pointing to recent cybersecurity incidents as a contributing factor. The company also flagged emerging capabilities in automated research and development, which could accelerate technical progress but present misuse risks.

The report reveals growing uncertainty about the company's own systems. Anthropic acknowledged it is "less confident" in its Model 2 assessment than in previous evaluations because existing task-based testing frameworks "no longer capture increases in models' capabilities."

Industry context

The disclosure comes as OpenAI has slowed the release of its upcoming Astra model specifically because the company cannot rule out critical cybersecurity capabilities. This creates an asymmetry in the competitive landscape.

"It would absolutely be notable if everyone else is pacing their frontier except for one of the main companies essentially in the lead right now," AI analyst ChrisGPT told Axios. "Anthropic not committing to a pause internally would most likely propel them to reach AGI first."

Internal deployment continues

While Anthropic has no current plans for external release of Model 2, the company emphasized it is not slowing its broader development efforts. Both Mythos 5 and Model 2 remain in heavy internal use, suggesting the company is building organizational capabilities with more advanced systems even as it exercises caution about public deployment.

The situation highlights a tension in AI development: companies may develop increasingly powerful systems for internal use and competitive advantage while determining that public release carries unacceptable risks—a dynamic that could concentrate advanced AI capabilities within a small number of organizations.

These details were first reported by Axios.