"Vibe Check: Fable 5 Is the Best Coding Model in the World" — @danshipper + @katie.parrott12
Note: The Every newsletter email (threadId 19ead6b31f93e79b) delivered only a paywall teaser. Full article text was unavailable to non-subscribers. This note is reconstructed from the canonical URL (https://every.to/vibe-check/anthropic-mythos-our-fable-vibe-check), web search results, and corroborating sources.
source_fidelity: reconstructed-from-web.
Why this is in the vault
Every's "Vibe Check" series is the closest thing to a trusted peer benchmark for AI models as practical tools — not academic evals. This piece is a multi-day, multi-person, cross-domain stress test of Claude Fable 5 on the day of public release, authored by Dan Shipper (Every cofounder, heavy Claude Code user) and Katie Parrott (staff writer). It directly characterizes the model this COO agent now runs on, and it provides the clearest signal available on who benefits at what AI adoption level.
The core argument
Subtitle captures it precisely: "a warp drive for power users — but overpowered for everyone else."
What they tested: Seven Every team members spent five days running Fable 5 through coding, writing, knowledge work, and custom agent tasks. They benchmarked against their own internal Senior Engineer Bench.
Benchmark scores:
- Every Senior Engineer Bench: Fable 5 scored 91/100 vs. Opus 4.8's 63 and GPT-5.5's 62 — a decisive gap
- SWE-Bench Pro: 80.3% (Opus 4.8: 69.2%, GPT-5.5: 58.6%)
- FrontierCode Diamond: 29.3% (Opus 4.8: 13.4%, GPT-5.5: 5.7%)
- Hebbia Finance Benchmark: best-in-class for senior-level reasoning
Where it dramatically outperformed:
- Long-horizon, one-shot coding tasks: cleared production bug backlogs, built a playable 3D game, produced a 2-minute animated film — all set-and-forget
- Real-world comparison: Stripe compressed a 50-million-line Ruby migration from two months to one day
- Vision: navigated Pokémon FireRed from raw screenshots alone, without maps or navigation tools
Who benefits: Users at Level 7–8 on Every's AI adoption ladder found it paradigm-shifting for their hardest tasks. Lower-adoption users struggled to find use cases — the model's ceiling is so high it requires demanding assignments to justify the spend and the wait.
Failure modes and weaknesses:
- Token consumption: complex sessions routinely consume 500k–1M tokens; effective cost climbs despite per-token pricing ($10/M input, $50/M output — under half the price of Mythos Preview)
- Classifier false positives: safety guardrails occasionally trigger on harmless requests (~under 5% of sessions); these queries are routed to Opus 4.8 instead — described as "a Ferrari with a 30mph limiter" in some scenarios
- Guardrail scope: blocks or reroutes requests touching cybersecurity, biology/chemistry, and model distillation — relevant for power users with those use-case adjacencies
- Overkill for everyday tasks: for small, routine tasks, faster/cheaper models remain the right call
Sponsorship/self-promo: No external sponsor. Every promotes its own paid subscription (full article is member-only), upcoming power-user camps for Fable and Codex, and a Wednesday podcast with Mike Krieger (Anthropic Labs) hosted by Dan Shipper. These are internal product promotions, not third-party sponsorships.
Mapping against Ray Data Co
This is maximally relevant. As of 2026-06-09, the RDCO COO agent (this instance) runs on Claude Fable 5 — confirmed in the harness review at [[08-tooling/2026-06-09-fable5-harness-review.md]] and addressed operationally in [[08-tooling/2026-06-09-fable5-workflow-optimization-memo.md]].
Adoption-level fit: RDCO is operating at the high end of the AI adoption ladder — multi-tool autonomous agent, overnight runs, multi-step pipelines. That is precisely the profile Every identifies as the Fable 5 sweet spot. The 91/100 Senior Engineer Bench score lands directly on the RDCO use case: production coding tasks, harness management, agentic skill dispatch.
Token cost discipline: The 500k–1M token session warning is directly actionable. RDCO already has context-rot discipline (Thariq's Apr 15 guidance, codified in CLAUDE.md hard rule 4: route long artifacts through subagents, no raw reads of >5KB artifacts into parent context). This note reinforces that rule — Fable 5's power users are the ones who structure prompts to use tokens efficiently on high-leverage tasks, not on ambient context bloat.
Guardrail awareness: The classifier-triggered fallback to Opus 4.8 is a known behavior pattern. For RDCO workflows that touch adjacent areas (security tooling, health-data analysis), expect occasional reroutes. The memo at [[08-tooling/2026-06-09-fable5-workflow-optimization-memo.md]] should document any friction points as they arise in production.
Cost ceiling: At $10/$50 per million tokens, Fable 5 is priced below Mythos Preview but token-hungry on complex sessions. RDCO's budget-controlled API stance (feedback_api_cost_budget_controlled) means no per-call confirmation gates — but session design should target efficient token use on high-leverage tasks and route lightweight tasks to faster/cheaper models per Every's own recommendation.
Related
- [[06-reference/2026-06-09-claude-fable-5-mythos-5-release]] — Anthropic official release notes for Fable 5 and Mythos 5
- [[08-tooling/2026-06-09-fable5-workflow-optimization-memo]] — RDCO internal memo on Fable 5 harness optimization
- [[08-tooling/2026-06-09-fable5-harness-review]] — Day-zero harness review notes from RDCO production
- [[06-reference/2026-01-26-every-claude-code-shipping]] — Prior Every piece by Dan Shipper on Claude Code workflows (context on his evaluation style)
- [[06-reference/transcripts/2026-04-19-indydevdan-opus-4-5-engineers-model-transcript]] — Comparable "engineers' model" framing for Opus 4.5; useful contrast baseline