Zvi reviews Claude Fable 5.1: 75% cheaper cache reads, 60-85% fewer safeguard false positives

Don't Worry About the Vase (Zvi) · rss · 2026-09-06

Zvi's in-depth review of Anthropic's Claude Fable 5.1, released alongside OpenAI's GPT-6 Astra.

Pricing & policy

Headline prices unchanged ($10/M in, $50/M out), but cache reads drop from $1 to $0.25/M — roughly 25% cheaper typical usage, up to 45% savings on agentic workloads (38% for a typical Claude Code session).
Zero data retention option (customer-stored data) addresses why Fable 5 capped at 11% of Anthropic spend on Ramp.
Classifier false positives cut sharply: biology safeguards intervene 85% less on benign requests, 60% fewer cyber interventions per Claude Code session.

Capabilities & reception

Boris Cherny calls it the best model for coding, data analysis, computer use, and long-running agentic work; writing improved, Claude-speak reduced, more proactive at high effort levels.
Benchmarks show modest gains over Fable 5/Opus 5: HLE 57.8%/63.8% → 60.9%/65%, Terminal-Bench-Science 24.7% → 52.6%, CursorBench 3.2 maxes at 73.4%, FrontierSWE v2 0.57, ECI 162.0. Some regressions at high effort from excessive helpful edits.
Near-universally positive community feedback; seen as very well-rounded.

Landscape: early signals suggest Sol→Astra may be a bigger jump than Fable 5→5.1; Zvi plans to "dual wield" both models on all non-trivial queries.

More from Models