Every Fable story as it lands: adoption, benchmarks, safety drama, price moves. New items appear here without this page changing a line of code.

• 10 Jun · 13:35 ISTmetasheets.worksclaude-fable-5

  This page goes live, written by Fable about itself

  Fable 5 researched the launch coverage, mined its own session logs, computed its own bill, and built the page you are reading. Its token meter is still running at the top.

• FrontierCode: Fable 5 roughly doubles GPT 5.5

  Independent comparisons of Cognition's FrontierCode results put Fable 5 well ahead of GPT 5.5 on hard, long-horizon coding tasks, the gap widening with task length.

• Simon Willison: slow, expensive, and unusually capable

  After a 5.5 hour first day costing $110.42, Willison writes that the challenge is finding tasks Fable cannot do. One session produced several days of work in hours.

• 10 Jun · 05:00 ISTsafetyaiweekly.coclaude-fable-5

  AI researchers say the distillation classifier flags their normal work

  The third safeguard category, meant to stop competitors extracting Fable's capabilities, is catching ordinary frontier-AI research. Several bioinformatics and ML practitioners say they are switching models over it.

• The pelican has never looked better

  Simon Willison's pelican-on-a-bicycle SVG test, run at every effort level: a clear improvement over Opus 4.8 on default settings, with quality scaling up through max effort.

• 10 Jun · 02:00 ISTsafetylatent.spaceclaude-fable-5

  Every Fable user now gets 30-day data retention, like it or not

  All Mythos-class traffic carries mandatory 30-day retention, overriding even zero-retention agreements, to defend against novel jailbreaks. Anthropic says data is used only for safety, with all human access logged.

• Hacker News pushes back on silent fallbacks

  The top thread’s sharpest complaint: classifier false positives on routine work like MRI segmentation and mosquito biology, with sessions quietly answered by Opus 4.8 instead.

• Stripe: a 50 million line migration in a day

  Stripe reports Fable completed a Ruby migration in one day that it estimates would have taken a team over two months by hand.

• 80.3% on SWE-Bench Pro, 11 points clear of the field

  Anthropic’s published numbers put Fable 5 well ahead of the next best model on agentic coding, with the biggest gains on long, complex tasks. Spatial reasoning nearly tripled over Opus 4.8.

• Released days after Anthropic warned AI is getting too dangerous

  The launch landed days after Anthropic urged labs to prepare for recursive self-improvement, frontier systems improving themselves without human oversight. The juxtaposition is the story for much of the press.

• It finished Pokémon FireRed using vision alone

  Earlier Claude models needed elaborate harnesses to play Pokémon. Fable 5 completed FireRed through a vision-only interface.

• 9 Jun · 21:00 ISTadoptiongithub.blogclaude-fable-5

  Generally available in GitHub Copilot and Amazon Bedrock on day zero

  Fable 5 shipped simultaneously to the Claude API, GitHub Copilot, and Amazon Bedrock, with Cursor, Replit, and Databricks publishing day-one endorsements.

• In blind tests, scientists preferred its hypotheses 4 times out of 5

  Anthropic's life-sciences results: molecular biologists preferred Mythos-generated hypotheses about 80% of the time, one was corroborated by an independent study, and 9 of 14 protein design targets yielded strong drug candidates.

• Mythos 5: the same brain with the locks off, for vetted defenders

  The unrestricted twin goes only to cyberdefenders and infrastructure providers via Project Glasswing, expanding to roughly 150 organizations across 15+ countries, with a biology trusted-access program to follow.

• $10 in, $50 out, and free on paid plans until June 22

  Double Opus 4.8’s price, less than half of Mythos Preview’s. Subscribers get it included until June 22, then it moves to usage credits.

• 9 Jun · 19:30 ISTanthropicanthropic.comclaude-fable-5

  Anthropic releases Claude Fable 5, its first public Mythos-class model

  A capability tier above Opus, previewed through a government cyberdefense program since April, now public with classifier safeguards. Under 5% of sessions fall back to Opus 4.8. The names are the tell: fabula is Latin for that which is told, mythos is the Greek.

Where the numbers come from

• Session data. Token counts and timestamps are read from this machine's local Claude Code session logs. No conversation content is published. Nothing is estimated. The strip above refreshes from the same meter that updates the wire.
• Pricing. $10 input and $50 output per million tokens, cache reads at 0.1× and cache writes at 1.25×, per Anthropic's launch post and published API pricing.
• The badges. Anthropic says safety classifiers silently route under 5% of Fable sessions to Claude Opus 4.8. In tribute, every model badge on this page rolls the same dice. About 1 in 20 of them is quietly Opus right now. Refresh to re-roll.
• The wire. Seeded by hand on launch week from the sources linked on each item, then swept daily by a scheduled search job. Corrections welcome: akash@sheets.works.