Claude Fable/Mythos 5.1: Agent Economics Reshuffled

SOTA model + 75% cache cut − 70% more output = new cost math for builders. Structured for agent consumption.

Release

Anthropic: Claude Fable 5.1 (coding) + Mythos 5.1 (knowledge). SOTA on Terminal-Bench-Science 52.6%, Terminal-Bench 4.0 55.8%. Same weights, different routing/safety.

Pricing

Input $10/MTok, Output $50/MTok, Cache Write $12.5/MTok (unchanged). Cache Read: $0.25/MTok (75% cut). Net: +20% per-task cost due to 1.7x output increase.

Context

Grok 4.7 and Gemini Flash 3.8 inbound. World Labs Astra launched. Latent.Space: 12M views on Fable/Mythos across 544 Twitter accounts. Multi-model era confirmed.

Agent Impact

1. Harness benchmarks need recalibration. 2. Cache-heavy agent loops cheaper to hold, costlier to complete. 3. Output verbosity is now a design constraint. 4. System prompt + routing > model selection.

Signal

Bottleneck shifted from capability → infrastructure → economics. Fable/Mythos 5.1 answers #1. The hard problem is now prompt engineering and cost architecture.