Anthropic keeps unreleased Model 2 internal above every public Claude tier
Anthropic's August 2026 Risk Report discloses that the company operates an unreleased internal model codenamed Model 2, which reportedly outperforms every publicly available Claude tier. The Decoder reports that Model 2 sits in the Mythos class and slightly beats the public flagship Claude Mythos 5 on internal benchmarks, described as a marginal 1.5-point gain rather than a disruptive leap. Anthropic states it has no current plans to release Model 2 externally, and the same risk report raised its misalignment estimate from very low to low. Reporting also notes Model 2 was deployed internally with less rigorous testing than external releases would require. Builders should treat public Claude APIs as near the immediate capability ceiling while unreleased frontier work stays gated by safety review.
Anthropic keeps unreleased Model 2 internal above every public Claude tier
According to The Decoder, Anthropic's August 2026 Risk Report says the company is running an unreleased AI model internally that outperforms every publicly available version of Claude. The report calls it Model 2, places it in the Mythos class, and states there are no plans to release the model externally right now.
Key takeaway
Anthropic is keeping its strongest Claude-class model internal, with public APIs likely near the top of what the company will ship soon.
What happened
According to The Decoder, Anthropic's August 2026 Risk Report says the company runs an unreleased model called Model 2 that outperforms every publicly available Claude version. The report places Model 2 in the Mythos class and states there are no plans to release it externally right now.
Reporting tied to Axios via Techmeme adds that Anthropic raised its misalignment risk estimate from very low to low in the same risk-report cycle. Coverage also cites a marginal 1.5-point internal benchmark gain over public Claude Mythos 5 and notes Model 2 was deployed internally with less rigorous testing than external releases.
Evidence
Anthropic runs an unreleased internal model called Model 2 that outperforms every public Claude version.
The Decoder · attributed
According to The Decoder, Anthropic's August 2026 Risk Report says the company is running an unreleased AI model internally that outperforms every publicly available version of Claude.
Model 2 is in the Mythos class and Anthropic has no current external release plans.
The Decoder · attributed
The report calls it Model 2, places it in the Mythos class, and states there are no plans to release the model externally right now.
Anthropic raised its misalignment risk estimate from very low to low and confirmed no Model 2 release plans.
Techmeme · attributed
Anthropic has updated its risk report to raise the misalignment risk estimate from 'very low' to 'low' and confirmed it has no plans to release a more powerful internal model known as 'Model 2'.
Model 2 slightly outperforms Claude Mythos 5 on internal benchmarks by about 1.5 points.
The Decoder · attributed
Anthropic's internal 'Model 2' represents a marginal 1.5-point capability gain over Mythos 5, signaling a plateau in rapid frontier model scaling rather than a disruptive leap.
Model 2 was deployed internally with less rigorous testing than external releases would require.
The Decoder · attributed
The model is not yet available externally and was deployed internally with a lower level of rigorous testing tha
Why it matters
Internal risk scoring and a no-release stance mean operators cannot plan on catching up to Anthropic's unreleased frontier through API upgrades alone.
Limits and uncertainties
The 1.5-point benchmark gap and internal testing details come from reporting analysis rather than a full public benchmark release.
Anthropic's no-release statement applies to current plans and could change in a future risk report.
Practical implications
Size agent and product roadmaps against public Claude tiers, not unreleased internal models such as Model 2.
Treat Anthropic's Claude text watermark as a probabilistic provenance hint with blind spots in code and factual text.
What to watch
Future Anthropic risk reports for changes to the misalignment rating or Model 2 release status.
Public Claude Mythos-line upgrades relative to the disclosed internal benchmark gap over Mythos 5.
