▶ YouTube Summaries

• https://www.youtube.com/watch?v=5LqC6qdVAwU
• Original title: You were lied to about Fable

Theo pushes back on the Twitter narrative that Fable 5 is nerfed, too expensive and blocked from coding. His claim: the model blew him away, most viral takes are wrong, and he wants to clear misconceptions before a follow-up video showing his actual workflow. Covers three things — the "nerf", subscription availability, and cost.

Performance is not nerfed

Anthropic's poorly-worded note ("some routine tasks like coding and debugging will fall back to Opus 4.8") spooked people. In 1.5 days of real work Theo hit zero fallbacks on normal tasks. Reroutes still fire on trigger words like cryptography/cyber (DEF CON-style PDF-decode puzzles get refused — but Opus 4.8 refuses those too). He hit one fallback configuring a dev Android phone with self-signing: mid-response it detected trigger words, nuked and rewrote the message as Opus 4.8, then went back to Fable after a couple exchanges.

The safety system is not keyword matching. It runs classifier models between request and response, screening both input AND output — so every request got more expensive, and Anthropic bills on response + input tokens, not on the safety check. Current design is a cheap two-stage system: stage one is a probe watching Claude's internal activations (very cheap); if sensitive sections trigger, it calls a more expensive classifier. This is likely why crypto puzzles get blocked — the region of the model needed to solve them overlaps with the region used to hack. History: constitutional classifiers cut jailbreak success 86% → 4.4% but added 23.7% compute; best-in-class added >50%; the new two-stage approach knocks harmless-query refusals to 0.05% at ~1% overhead. Takeaway: they're evolving it, it'll have bugs, but this is the wrong thing to attack Anthropic over. Blanket category bans (e.g. sandbagging ML tasks) deserve complaint; this doesn't.

The viral "model got dumber" benchmark is nonsense — an unreliable, noisy bench from someone who has posted wrong numbers before, testing lots of terms the model currently dislikes. Its reasoning leaderboard ranks Sonnet 5, GLM 5.2, NeMoTron 3 Ultra and others above Opus 4.8, which is absurd. No one else reports the regression.

Subscription availability is a deliberate experiment

Fable now has its own weekly limit, capped at half your total weekly limit — Anthropic saw from the earlier 3-day window how GPU-heavy it is and won't starve enterprise customers. Theo only hit 23% in a heavy day. Fable stays in subs through July 7th, then moves to usage credits. This is NOT a plan to upsell individuals a $1000 tier — Anthropic's money is in enterprise. The real constraint is compute: Fable only exists in subs at all because of the xAI/Colossus deal for GPUs. It launched at less than half its originally-planned output price (~125 → 50/M) because they had enough chips. The 7-day window is marketing hype plus user research — a full week of power-user weekday/weekend data to size GPU allocation before enterprises ramp up and availability plummets. Anthropic stated it will restore Fable to subs "as soon as capacity allows".

Cost control

• Never use effort above high. X-high and Max second-guess and loop, giving worse over-done code at 10–50× cost with basically zero quality gain. Ultra-code just spawns many high-effort runs. Stick to high; try low/medium too.
• Fable selected doesn't mean Fable does all the tokens. Theo taught Claude Code to shell out to Codex/GPT-5.5 — generous Codex limits, better at computer use and vision, and good for token-hungry work (PDFs, auditing large codebases, screenshots). Fable excels at orchestrating a fleet of sub-agents.
• One thread triaged ~25 open PRs (merged/rewrote/closed) in ~5 hours for ~$150–200 of notional usage across all models — easily inside two $200 subs (Codex ~10–15%, Claude ~40%). Also runs a "vibe proxy" to split traffic across accounts.

Recap: use sub-agents/cheaper models for cheap tasks, never touch X-high/Max/Ultra, and treat the current sub window as an experiment. The nerf story is overblown; Anthropic's fault is bad comms, not a bad model.