Anthropic Unleashes Mythic‑Level Claude 5 and Claude Fable 5 – A Massive Performance Leap

Anthropic announced the simultaneous launch of Claude Fable 5 and Claude Mythos 5, describing the latter as a "mythic‑level" model that had been hidden for two months due to safety concerns. Both models share the same underlying architecture, but Fable 5 includes an internal safety classifier that degrades responses to Opus 4.8 when a security‑related request is detected.

The pricing for Fable 5 matches Opus 4.8 Fast Mode at $10 per million input tokens and $50 per million output tokens, making it roughly twice the cost of Opus 4.8 but far cheaper than Mythos Preview or GPT‑5.5 Pro.

Benchmark dominance : On the SWE‑Bench Pro coding benchmark, Fable 5 scores 80.3%, outpacing Opus 4.8 (69.2%), GPT‑5.5 (58.6%) and Gemini 3.1 Pro (54.2%). FrontierCode Diamond shows a five‑fold lead (29.3% vs. 5.7%). In agent‑centric tasks, Fable 5 achieves three times the success rate of Opus 4.8 when playing "Slay the Spire" with persistent memory.

Real‑world demos : The models autonomously completed a 50‑million‑line Ruby code migration for Stripe in a single day, built a fully functional Factorio factory from scratch, generated a complete 3D printable CAD editor and model, and played through Pokémon Red using only screen‑capture inputs—no external tooling or game state information.

Scientific breakthroughs : Mythos 5 performed end‑to‑end protein design, delivering nine strong candidate solutions across 14 disease targets, and independently gathered data from 138 species to train a genomics model that outperforms recent Science‑published work despite being 100× smaller.

Safety and jailbreak resistance : Anthropic added a safety layer that routes network‑security, bio‑chemical, or model‑distillation queries to Opus 4.8. Red‑team testing (400 rounds) shows that only about 0.03% of traffic is affected, though some benign tasks still trigger the safety downgrade.

Overall, the release positions Claude Fable 5 and Mythos 5 as the current state‑of‑the‑art LLMs across coding, multimodal reasoning, and scientific research, while highlighting the trade‑off between unrestricted capability and safety controls.