▶ YouTube Summaries

• https://www.youtube.com/watch?v=2kObwHxHILQ
• Original title: Coding with Fable 5 is actually insane.

Hands-on review of Anthropic's Claude Fable 5, framed as the publicly available, heavily safeguarded version of the Mythos-class model. Forrest tests it on security audits, feature work in TypeScript and Rust codebases, and a one-shot app build from a spec file. Verdict: the best coding model he has used, especially for goal-driven autonomous work with parallel agents, but it is roughly 10x slower and 5x more expensive per task than Opus 4.8 / GPT 5.5, and its trigger-happy safeguards (auto-switching to Opus 4.8 on anything touching cybersecurity, biology, or ML research) make for a frustrating user experience. Good model, bad company economics.

What Fable 5 is

• A Mythos-class model "made safe for general use"; Anthropic explicitly says it is released instead of Mythos because it carries safeguards.
• Mythos 5 (the unrestricted upgrade of the Mythos preview) is reserved for a small group of cyber defenders and infrastructure providers — not available to the public.
• Pricing: $10 per 1M input tokens, $50 per 1M output — 2x Opus 4.8. On subscription plans it burns limits 2x faster; Forrest exhausted his $100/mo 5x Max plan in about 30 minutes.
• After June 23 Fable 5 is removed from Pro/Max/Team/Enterprise plans and becomes API-only via usage credits — so no more subsidized access; he warns against becoming dependent on it.

Safeguards

• Cybersecurity and biology queries get flagged visibly and rerouted to Opus 4.8. ML research / LLM-development safeguards are invisible: instead of a model switch, effectiveness is throttled via prompt modification, steering vectors, or parameter-efficient fine-tuning.
• Extremely trigger-happy: it refused to answer "is the mitochondria the powerhouse of the cell", refused a question about what security research it is allowed to do, refused to write up a report about a security audit it had itself just performed, and refused to scan Hyprland's C++ codebase with the same prompt that worked on his own repo. Even mentioning "runtime" mid-session tripped the filter once.
• Karpathy (now at Anthropic) admits the safeguards are "too trigger happy for launch" and should be tuned so adjacent topics stop flagging — but core cyber/bio blocks will stay.

Enterprise data-retention gotcha

• Fable 5, Mythos 5, and future Mythos-class models require 30-day data retention on all traffic, first and third party (including GovCloud/FedRAMP/BAA routes). Zero-data-retention customers cannot use them. Anthropic says data is safety-use only, human access logged, deleted after 30 days "in almost all cases" — Forrest notes that "almost" is doing a lot of work.

Benchmarks and vision

• Announcement benchmarks group Mythos 5 + Fable 5 in one column; best in nearly every category, but Mythos 5 carries the cyber/bio numbers since Fable 5 simply cannot do those tasks (effectively 0%). He distrusts coding benchmarks generally.
• Vision highlight: Fable 5 beat Pokemon Fire Red start-to-finish from raw screenshots with only a minimal vision-only harness (previous Claude models needed heavy helper harnesses and still struggled). It can also rebuild a web app's source from screenshots.

Security audit of his URL shortener (TypeScript / TanStack Start)

• Deployed 6 parallel review agents to scan for vulnerabilities, then double-checked and remediated. Five findings:

1. Real vuln: a public server-function endpoint (every createServerFn is a publicly callable RPC in TanStack Start) returned the full link row including bcrypt password hash, user ID, workspace ID — no auth. It was dead code; Fable 5 correctly just deleted it.
2. False positive: flagged "privilege escalation" where admins can invite admins — but role gating exists (ensureWorkspaceRoleAssignable) and the behavior is the documented, intended feature. It missed the gating code and mistook a feature for a bug; on double-check it discarded the finding itself. Lesson: never ship AI findings without reading the code yourself.
3. Real and worse than first reported: reserved-slug check (signin/signup/dashboard) was case-sensitive and duplicated in a second fallback route missing six more routes; the double-check pass found the second path and fixed both.
4. Fix quality mixed: created a shared escapeHtml util but left the duplicate local copy; fixed a real dashboard bug (transient errors locked the page behind the paywall) but discriminated errors by regex on human-readable text instead of a typed server error code.

Head-to-head cost/speed

• Same feature, three models: Opus 4.8 and GPT 5.5 each took 2–3 minutes and ~$2–3; Fable 5 took 23 minutes and ~$11. All three produced passable, working code. So: 10x slower, 5x pricier for everyday tasks.

Where it actually shines

• Goal-driven, long-running autonomous work. Give it one goal or a spec and it orchestrates parallel agents (8 for a full app redesign) unlike anything he has seen. This matches Boris Cherny's "work in loops, don't prompt" workflow — which makes sense given Anthropic used Mythos-class models internally.
• Demo: he co-wrote a 225-line design doc (vision, scope, phases, build order, failure modes) for a "soundscape time machine" — pick a city street, drag a timeline 1850–2026, hear synthesized street audio generated purely from math in the browser, no recordings — gave Fable 5 the MD file with "do it, don't stop until done", and it built the whole thing.
• Caveat: such runs produce 500k+ lines of code to review in ~30 minutes, contradicting the "read the code yourself" lesson from the audit.

Takeaway

• Best coding model he's used on high effort, but for everyday back-and-forth coding the speed/cost trade-off doesn't pay; keep it in your back pocket for tasks other models fail at, especially once it goes API-only. Broader unease: Anthropic reserving true frontier models for a few hundred approved companies while the public gets the safeguarded tier — "good model, bad company."