Claude Fable 5.1 Lands With 75% Cheaper Caching, as Meta Cuts Token Prices 90%

Ben's Bites · rss · 2026-09-03

Highlights from this Ben's Bites issue:

Claude Fable 5.1 is out: no dramatic quality jump, but faster and easier to talk to per the author; a new system prompt bans repeating song lyrics and copyrighted logos. Anthropic also cut input caching costs by 75%, making API usage 25% cheaper overall.
New model price war: Gemini 3.8 Flash and Meta's Muse Spark 1.3 match on price and performance; Flash is 50% off for the year while Meta offers a 90% cut in exchange for data. Meta's Muse Voice Transcribe looks extremely strong for transcription.
Gemini video understanding: instead of fixed-rate frame sampling, it searches frames, audio and transcripts for relevant moments — 88% fewer tokens, up to 66% cheaper, with a small accuracy bump.
Tools & ecosystem: Claude Code/Cowork can now use your computer in the background; WebMCP lets agents call site actions directly (Vercel's agent-browser supports it); Shopify beat GPT-5.6 Sol xhigh on a narrow task with a 0.8B model and open-sourced the infra as Tangle; Stanford scrapped 85% of its software course in favor of agents, taste, and real OSS PRs.
Author builds: Ben shipped a real-time token price tracker via OpenRouter and one-shot a remove.bg replacement after Canva shut it down, all open-source — and ran out of his weekly Fable quota.

More from Models