Anthropic's Internal "Model 2" Leaked, Outperforms Mythos 5
koltregaskes · x · 2026-08-17
Anthropic's new Risk Report mentions an unreleased internal model referred to as "Model 2".
Key Facts:
- Capability: Described as "somewhat more capable than Mythos 5" and noticeably better than MythoMax and 4.0-Snapshot.
- Benchmark: Achieves a score of 86.7% on MMLU (5-shot, MMLU-Pro).
- Status: One of Anthropic's most capable and commonly used internal models, with no current plans for external release.
- Speculation: Model 2 is likely a fine-tuned or specific deployment of Claude 3.6 Sonnet, or an intermediate product of continuous training on a new architecture.
More from Companies & People
- Claude Desktop adding built-in browser and model comparison mode — testingcatalog · 2026-08-17
- AI Safety Researcher Seeks Funding for Independent Work — JacquesThibs · 2026-08-17
- Gary Marcus Criticizes the Hype Around Anthropic's IPO — Gary Marcus · 2026-08-17
- Satirizing AI marketing: Selling AI to protect against AI displacement? — suchenzang · 2026-08-17
- Writing drives understanding: Stop hesitating and start writing — tokenbender · 2026-08-17
- Dario Amodei responds to criticism of AI messaging, claiming balance — iamKierraD · 2026-08-17