Anthropic Launches Claude Fable/Mythos 5.1: New SOTA with 20% Higher Task Cost
Latent Space · rss · 2026-09-02
Anthropic released Claude Fable 5.1 (coding/long-horizon tasks) and Mythos 5.1 (knowledge work), claiming they are the world's most advanced models.
Key Updates & Metrics
- Performance: Scores 66 on the Artificial Analysis Intelligence Index, surpassing Opus 5 (63) and Fable 5 (62). Terminal-Bench Science more than doubled (24.7% → 52.6%).
- Pricing: Input/Output/Cache Write prices remain ($10/$50/$12.5), but cache read price slashed by 75% to $0.25/MTok, significantly benefiting agentic workloads.
- Cost Trade-off: Despite cheaper cache, the model uses 1.7x more output tokens, leading to a net 20% increase in cost per task ($3.14 → $3.76). GPT-5.6 Sol remains the value leader.
Technical Details & Controversies
- Architecture Speculation: Community analysis suggests Fable 5.1 and Mythos 5.1 may share the exact same underlying weights, differing only in safety/routing behavior rather than being distinct base models.
- Fallback Mechanism: Benchmarks revealed that 4% of output tokens were routed to Opus 4.8/5 due to safety triggers, raising questions about "true" capability.
- Agent Positioning: Official messaging emphasizes autonomous, multi-step workflows and introduces Enterprise Frontier Safeguards (EFS).
More from coding & agent
- After LLMs refused to crack DRM, an agent scripted screenshot+OCR to translate books — evisoft · 2026-09-02
- Fable 5.1 builds a macOS animation app in under 20 minutes — amos_gyamfi · 2026-09-02
- Testing AI Agents: Complement Not Replacement for Scripts — PartyVermicelli1870 · 2026-09-02
- Experimenting with Claude to Revive streamlit-drawable-canvas — andfanilo · 2026-09-02
- Would developers pay for an 'infrastructure layer' that equips AI agents with phone, email, and memory? — Athlore_AI · 2026-09-02
- Dev notes record outcomes, but the reasoning dies in the transcript — Sea-Perception1619 · 2026-09-02