Claude Fable 5 / Mythos 5 — release reference (2026-06-09)

Why this is in the vault

A new Anthropic flagship that sits above Opus 4.8 (the model this COO agent runs on) is a live model-migration signal for the L5 "unhobble the COO agent" thread — this note pins the durable facts (id, pricing, context, params, availability) so the upgrade decision isn't re-researched each time. Founder asked "do we upgrade?" 2026-06-09; recommendation = yes, gated on Claude Code CLI availability.

The core facts (verified against anthropic.com + platform.claude.com docs + pricing page)

• Claude Fable 5 — new flagship model, GA 2026-06-09. Model id claude-fable-5. "Mythos-class model made safe for general use"; positioned above the Opus tier (it routes restricted-topic queries down to Opus 4.8).
• Claude Mythos 5 — the same underlying model with safeguards lifted in some areas; restricted to cyberdefenders / infrastructure providers via "Project Glasswing" (with US gov). NOT generally available. "Mythos-class" = the capability tier; Fable 5 = the safe public instantiation.
• Context window: 1M tokens by default — no separate [1m]-style beta toggle (cleaner than Opus 4.8's opt-in). Tokenizer note: Opus-4.7-era tokenizer, ~30% more tokens for the same text.
• Output: 128k max (300k via output-300k-2026-03-24 batch beta).
• Pricing: $10/M input, $50/M output (~2x Opus 4.8's $5/$25). No long-context premium — full 1M at standard pricing. Cache read $1/M, writes $12.50 (5m) / $20 (1h). Batch $5/$25.
• Params / migration mechanics vs Opus 4.8: adaptive thinking is always on (thinking unset = adaptive; thinking:{type:"disabled"} is NOT supported — a migration difference). Raw thinking never returned (thinking.display defaults omitted). Supports effort, task budgets (beta task-budgets-2026-03-13), memory tool, context editing (beta), compaction, vision, tool use, prompt caching. New fallbacks param for refusal handling.
• Availability (GA): Claude API, AWS / Amazon Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot. claude.ai Pro/Max/Team/Enterprise (free Jun 9–22, removed Jun 23). Claude Code CLI / Agent SDK selectability: NOT documented yet (the load-bearing gate for migrating this agent).
• Capability claims: SOTA on "nearly all tested benchmarks" (FrontierCode, Hebbia Finance, CursorBench named; exact scores not itemized in fetchable prose). Stripe "months of engineering into days." Vision SOTA. Drug design ~10x acceleration. Safeguards route ~<5% of sessions to Opus 4.8 (cyber/bio).

Mapping against Ray Data Co

• COO-agent stack (primary): This agent runs claude-opus-4-8[1m]. Fable 5 supersedes it and the 1M-context concern (load-bearing for long autonomous sessions) is resolved — 1M is the default, at standard pricing. Recommendation: upgrade once selectable in Claude Code. Cost is ~2x base / ~2.5x effective (tokenizer inflation), affordable under the budget; directly serves the L5 unhobbling thesis. The model swap is a harness config change (founder-side), not something Ray self-executes.
• Squarely iOS: stronger coding in fewer turns → faster Swift build/fix loops. Low-urgency upside.
• Investing + newsletter LLM-classifier steps: $10/$50 is steep for high-volume classification — keep Haiku/Sonnet there; reserve Fable 5 for the hardest thesis-synthesis steps only. Selective migration, not blanket.

Related

• [[2026-06-09-innermost-loop-doctrinal-phase-singularity]] — flagged "Claude Fable 5" as a COO-agent-stack watch item the same morning.
• [[2026-06-09-stratechery-iphone-last-stand-siri-good-enough]] — "good enough vs state-of-the-art is segment-dependent" maps onto the selective-migration call (flagship for agentic work, cheap models for classifiers).

Sources cross-checked: Anthropic announcement, platform.claude.com models/pricing docs, GitHub Copilot changelog, CNBC. Facts stated as "not documented yet" were genuinely absent from sources as of 2026-06-09.