Top story
Claude Fable 5.1 cuts the part of the bill agents keep rereading
https://www.anthropic.com/claude-fable-and-mythos-5-1
Claude Fable 5.1 keeps input and output pricing at $10 and $50 per million tokens, but cuts cache reads by 75% to $0.25. Anthropic estimates roughly 25% lower cost for typical workloads and as much as 45% for highly agentic ones. It also reports Terminal-Bench-Science rising from 24.7% to 52.6%, although that remains a vendor-run evaluation with a 3.5–4.5 point standard error. simonw isolates the more decision-useful claim: later calls in a long-running agent can reuse cached transcript, so this discount reaches the repeated context directly. Fable is generally available; Mythos is the same model with more permissive safeguards for vetted cyber and life-science users. I see a release worth replaying against real traces, not a benchmark table to accept as neutral. Teams already routing Claude should compare identical tasks, effort settings, cache-hit rates, and intervention counts before changing their default.
Read the benchmark at its actual scope
A small transformer reaches 44% on ARC-1 for 67 cents. The write-up trains from scratch in 1.5 hours on an RTX 5090 and also scores 7% on ARC-2. Its ablations credit 3D RoPE and per-task embeddings for most of the gain. I keep embedding-shape’s caveat: this is an efficient ARC specialist, not evidence of a general model beating LLMs.
A tool to try with verification attached
Keenable SELECT turns web search into SQL. Its showcase exposes one MCP tool that runs read-only DuckDB queries, applies exact filters before semantic operators, and stores result sets plus the full trajectory. I like the audit trail and the chance to reduce token-heavy page reading. The semantic functions still make model judgments, so every report needs source links and an independent verification pass.
Cheaper context, smaller models, and clearer queries matter when they make evidence easier to inspect.
— Tin