Initial impressions of Claude Fable 5 (Simon Willison)

Simon Willison's hands-on review of claude-fable-5, anthropic's frontier model released 9 June 2026 — the same author and same shape as claude-opus-4-8-review, so it extends this wiki's model-substrate thread directly.

Key claims from the post

• Identity: Fable 5 "matches Claude Mythos 5's capabilities but with much more strict guardrails" — i.e. a safety-hardened sibling of a parallel "Mythos 5" release (the source's framing of a capability-equal, more-restricted variant).
• Specs: 1M-token context, 128K max output, Jan 2026 knowledge cutoff.
• Pricing: $10 / $50 per Mtok input/output — double claude-opus-4-8's $5/$25.
• Impressions: a "beast" — notably slower and more expensive, but handles complex work; "substantially larger" than predecessors on the evidence of speed, price, and knowledge depth.
• Knowledge gain: lists more of Willison's own open-source projects, with accurate dates — a concrete knowledge-breadth improvement over claude-opus-4-8.
• Agentic/coding feats in testing: converting a MicroPython sandbox to a full CPython-WASM build; advanced tool-call features for Datasette Agent; shipping a production Python wheel (a $110.42 full-day usage cost — the steep-price point made concrete).
• Availability: free on the Max subscription through 22 June 2026, then metered.

Caveats

Single practitioner hands-on (no benchmarks table here), post-dates this wiki's knowledge cutoff, and pricing/availability are dated snapshots. The "guardrailed-sibling" framing is now confirmed by the primary Anthropic announcement (same model, safety classifiers the only difference) — see that page and claude-mythos-5.

Tier: T4 → T2 (2026-08-25)

The old grade read the publisher: a personal blog. The page rests on what Willison did with the model — a full day's work costing $110.42, a MicroPython sandbox converted to a CPython-WASM build, and a check of whether the model knew his own projects and their dates. That is somebody testing a thing they neither built nor sell, which is T2. The specs and pricing in the list above are Anthropic's, relayed, and the primary announcement now carries them; they are not what this page is for. Still a single practitioner with no benchmark table, and that caveat stands where it always did.

Related

claude-fable-5 · claude-opus-4-8 · claude-opus-4-8-review · anthropic · synthesis