Anthropic's Internal Model 2 Capabilities Revealed

Anthropic publicly revealed an internal model codenamed 'Model 2' in its second Risk Report, classified as 'Mythos level' and slightly superior to Mythos 5. The model scored 62.8% on CoBench v2, a benchmark testing historical AI research task capabilities, surpassing Mythos 5's 50.3%. It is currently extensively used for coding, data generation, research, and agent tasks internally but has not completed full pre-deployment assessments. Anthropic stated there are no plans to release Model 2 publicly, without providing specific reasons.

Confirmed

• Mythos 5 and Model 2 are Anthropic's most capable and frequently used internal models; Model 2 is slightly better than Mythos 5 and is restricted to internal use
• Model 2 scored 62.8% on CoBench v2; Anthropic estimates a score of 85% is required to fully replace its researchers, indicating a current gap
• Model 2 has incorporated anti-distillation as a regular engineering effort
• The report did not disclose Model 2's parameters, architectural details, or reasons for not releasing it

Unconfirmed

• @bindureddy claimed the non-release is due to 'alignment' concerns, noting its Fable 5 version remains fragile and prone to random hypotheses; he argued this concentrates superintelligence in one company, forcing others to catch up
• @haider1 speculated that compute is the main bottleneck, making public release unfeasible, hence limiting it to internal use or a select few companies, suggesting both OpenAI and Anthropic hold better internal models
• @scaling01 estimated Model 2 is about 1.5 points higher than Mythos 5 (approx. 162.79 points) and may have finished training in May-June; also speculated Mythos Preview is a 10T parameter 'teacher' model, with Mythos 5 and Fable 5 as distilled versions
• @imjustnewatai guessed the public might never access this model, creating a 'permanent underclass of intelligence'

Why it matters

• The 85% threshold provides a quantitative metric for 'AI replacing AI researchers': 62.8% means the model can handle a significant portion of historical research tasks, but a clear gap remains before full replacement
• 'Withholding the strongest internal model' has sparked discussions on the disparity between public and private frontier capabilities: comments in @teortaxesTex's thread suggest the strategy might change after Astra's release; @bindureddy views the delay as a window for the open-source community to catch up

2026-08-15 ~ 2026-08-16 · 21 related posts

• Anthropic confirms Mythos 5 as top internal model, hints at mysterious Model 2 — scaling01 · 2026-08-15
• Rumor: Anthropic uses internal model better than Mythos 5 — kimmonismus · 2026-08-15
• Rumor: Mythos Preview may be a 10T parameter teacher model — scaling01 · 2026-08-15
• Anthropic Report Leaks 'Model 2', Slightly Stronger Than Claude 3.5 Sonnet — ChrisGPT · 2026-08-15
• Anthropic Says It Has No Plans to Release Its Most Powerful Internal Model — teortaxesTex · 2026-08-15
• Leak: Anthropic's Internal 'Model 2' Beats Mythos 5 by 12.5 Points on CoBench, Approaching Researcher-Level — daniel_mac8 · 2026-08-15
• Anthropic's internal model scores leak: new model may surpass Mythos 5 — scaling01 · 2026-08-15
• [source] Report: Anthropic's Internal 'Model 2' Outperforms Mythos 5 on Diagnostics — imjustnewatai · 2026-08-15
• [source] Rumor: Anthropic Uses Internal Model Far Better Than Mythos 5 — Neurogence · 2026-08-15
• Anthropic Reveals Unreleased Internal Model That Beats Mythos 5 in 186-Page Risk Report — 机器之心 · 2026-08-15
• Anthropic's hidden internal model sparks talk of a permanent intelligence underclass — imjustnewatai · 2026-08-15
• Anthropic reportedly halts Mythos-2 release over alignment concerns — bindureddy · 2026-08-15
• Anthropic's Risk Report details internal Model 2 — still no substitute for its researchers — AccBalanced · 2026-08-15
• Speculation suggests Anthropic's internal model progress may be slowing — haider1 · 2026-08-15
• Anthropic Delays Model 2 Release Due to Compute Constraints — haider1 · 2026-08-16
• Leak: Anthropic's Internal Model 2 Outperforms, RSI Accelerating — teortaxesTex · 2026-08-16
• Major Labs Achieve RSI: OpenAI, Deepseek, and the Race to AGI — teortaxesTex · 2026-08-16
• Bindu Reddy: Anthropic Concentrates Power by Not Releasing Mythos 2 — bindureddy · 2026-08-16
• Anthropic accused of concentrating power by withholding Mythos 2 — bindureddy · 2026-08-16
• Anthropic report: Anti-distillation is now standard engineering — aigclink · 2026-08-16

1 near-duplicate retellings: kimmonismus