Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% Cheaper Cache Reads
Anthropic released Claude Fable 5.1 (general availability) and Claude Mythos 5.1 (restricted), sharing the same underlying model with different safeguard layers Terminal-Bench-Science 0.1 score of 52.6% roughly doubles Fable 5's 24.7%, outperforming Opus 5 (29.0%) and GPT-5.6 Sol (22.4%) Cache read pricing drops 75% from $1.00 to $0.25 per million tokens, yielding ~25% cost reduction on typical workloads and up to 45% on agentic ones Three breaking API changes: forced tool use removed, thinking
Analysis
TL;DR
- Anthropic released Claude Fable 5.1 (general availability) and Claude Mythos 5.1 (restricted), sharing the same underlying model with different safeguard layers
- Terminal-Bench-Science 0.1 score of 52.6% roughly doubles Fable 5's 24.7%, outperforming Opus 5 (29.0%) and GPT-5.6 Sol (22.4%)
- Cache read pricing drops 75% from $1.00 to $0.25 per million tokens, yielding ~25% cost reduction on typical workloads and up to 45% on agentic ones
- Three breaking API changes: forced tool use removed, thinking blocks are model-bound, and editing earlier turns now invalidates thinking blocks
- Cyber safeguards relaxed for vulnerability discovery (60% fewer interventions in Claude Code); biology safeguards fire 85% less on benign requests
Why It Matters
This release signals Anthropic's push into agentic scientific research workflows, where the 52.6% Terminal-Bench-Science score demonstrates a meaningful leap in autonomous reasoning over complex tool-using tasks. The aggressive cache pricing cut directly addresses one of the biggest cost barriers for production agentic deployments, making long-context agent loops economically viable at scale.
Technical Details
- Architecture & Specs: Both models share identical underlying weights with 1M token context window and 128K max output tokens; adaptive thinking is always enabled
- Benchmark Performance: Terminal-Bench 4.0 scores 55.8% (Fable) vs 60.9% (Mythos), revealing the performance cost of safeguard interventions; CursorBench 3.2.0 at 73.4%, OSWorld 2.0 at 41.7% strict
- Pricing: Base input $10/M tokens, output $50/M tokens unchanged; cache reads dropped to $0.25/M (0.025× base vs 0.1× on other Claude models); batch processing at $5/$25 per million
- API Breaking Changes:
tool_choiceset toanyortoolnow returns 400 (useautowith strict tool use); thinking blocks are model-bound so router/fallback chains lose reasoning when switching down; mid-conversation system/tools array edits error on accounts created after August 31, 2026 - Safeguard & Research Impact: Mythos 5.1 designed protein binders with ~50% hit rate across 12 targets vs 10-15% baseline; Fable 5.1 generated Venus elevation maps at 2-3 km resolution; custom GPU kernels accelerated seven open-source genomics models up to 2.5×
Industry Insight
- Agent architectures relying on conversation history manipulation (injecting reminders, rebuilding system prompts mid-turn) will need immediate refactoring; turn-scoped system messages and server-side context editing are the new patterns to adopt
- The cache pricing shift makes long-context agentic loops significantly more economical—teams running heavy tool-use workflows should re-evaluate cost models and may see 45% reductions, potentially unlocking previously uneconomical use cases
- The deliberate transparency around the Fable/Mythos performance gap (55.8% vs 60.9% on identical weights) sets a new industry standard for honest benchmarking; expect competitors to face pressure to disclose similar safeguard overhead figures
Disclaimer: The above content is generated by AI and is for reference only.