▶ YouTube Summaries

• https://www.youtube.com/watch?v=7IewbRdaBWI
• Original title: Fable is Mythos, and it is really good.

After ~24h and ~$2-3k of inference, Theo reviews Anthropic's new Mythos 5 model series. Cloud Code users actually get Fable 5 — Mythos with safety guardrails bolted on front. His verdict: the best coding model ever released by a wide margin, but brutally expensive, limit-burning, and at times silently dumbed-down by safety interventions.

What you actually get: Fable 5, not Mythos 5

• Mythos 5 is the raw model; Cloud Code users get Fable 5 — same model with safety guardrails in front.
• Guardrails cause refusals and routing that hurt benchmark scores (e.g. Terminal Bench dropped ~20 points vs raw Mythos because it refused some problems).
• Cloud Code subs only get Mythos until June 22; Fable included "for now", changes June 23, may extend "if capacity allows".

Benchmarks

• Crushes SWE-bench Pro (80% vs GPT 5.5's 58.6) — but Theo distrusts that bench (mostly recreating existing PRs from training data).
• Frontier CodeBench diamond tier: 30% vs Opus 4.8's 13% and 5.5's 5.7%.
• Vision: first Anthropic model to lead GPT; good spatial reasoning — but Google Gemini 3.1 Pro still dominates Theo's private Skatebench (98% vs Fable's 79%).
• Suspicious Frontier Code reasoning curve (scores go up/down randomly across reasoning levels — "looks like a random number generator").
• DeepSWE (which Theo trusts): Fable on x-high ≈ GPT 5.5, but using far fewer dollars — beats Opus scores for less spend despite double price.
• Artificial Analysis: now smartest model ever measured, +5 points over 5.5. Bombs HLE (refuses too much) while raw Mythos scores SOTA.

Pricing and limits — the big downside

• $10/M input, $50/M output — double Opus per-token.
• Burns limits fast: maxed two $200 accounts' 5-hour session limits in one workflow run; usage-based billing spent $100 in ~8 minutes.
• One 5-hour session ≈ 25% of weekly limit → ~4 full sessions/week.
• Expects a higher subsidized tier later given good margins + new GPUs ("Thank you, Elon").

Safety throttling — quietly making the model dumber

• Sensitive topics (cybersecurity, bio, chem) → transparently routed to Opus.
• But frontier LLM-development queries get silently degraded via prompt modification, steering vectors, PEFT — no notification, full price for a dumber model. Anthropic estimates ~0.03% of traffic affected.
• Artificial Analysis saw 8% fallback to Opus (Anthropic claimed <5%).
• Cloud Code refusals rare for Theo (~1 total); web app worse. Plans a dedicated video on this.

Real-world coding experience

• Feels like "more Opus turned up to 12" — thorough, harder-working, smarter, more willing to cheat.
• Ported his 2021 ping.gg app: 15k-line change, didn't one-shot but worked after ~4 error-screenshot rounds. Full stack overhaul (TanStack Start, Convex, Clerk) got close but broke core functionality.
• Better UI/design taste than Opus; less aggressively "AI-generated".
• Confidently confused itself on a staging-vs-prod audit, kept insisting a non-bug needed fixing even after correction.

What his team built in a day

• Terminal-based 2.5D text adventure (Rust), T3 Code ported to a Rust terminal UI, a full Minecraft clone with programmatically-generated textures (no image gen), and a multiplayer 3D racing game with spectator mode.

Data retention caveat

• Fable 5 requires 30-day retention for ALL traffic — even trusted/no-save setups. Anthropic says no training use, but data is stored → blocks many real-world/legal use cases. Likely distilling on Cloud Code histories.

How to use it / takeaway

• Treat it like a "laid-back senior engineer": give vague instructions, let it explore, write fuzzers, self-validate. Push for problems 5x harder, not 5% harder.
• Cleaned up stale PRs, rewrote data layers, modernized old codebases in hours. Strategy: $200 on 10 PRs (merge 3) beats $1000 on one unmergeable PR.
• Despite being an "anthropic-hater", says this is worth attention — the economics of software development changed fundamentally. Squeeze the subsidized window now.