▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=yh5IAUow4Hc
• Original title :: Nobody Is Competing With Anthropic

Theo reviews Anthropic's Fable / Mythos 5.2 after a day of use (~20 PRs landed). It tops most official benchmarks and is cheaper ($8/$40 per Mtok) and ~30% faster than 5.1, but in his own codebase-analysis evals it trails Opus 5.5 and other models. His verdict: Fable 5.2 is a brilliant but "spiky" specialist that needs more hand-holding, while Opus 5.5 remains the best everyday model. Anthropic dominates both the top-end and the workhorse tier.

The release

• Fable and Mythos 5.2 are the same model; Mythos just has fewer safeguards (cyber/bio) and is gated behind verification programs. Theo's analogy: one restaurant, two entrances with different guards.
• Claims: infeasible software migrations, three electronic devices designed from scratch, two open black-hole physics problems solved, Fermat's Last Theorem formalized in 30 hours, big robotics gains.
• Up to 30% faster output, cheaper ($8 in / $40 out vs $10/$50), new optional fast mode (2x price for 2.5x speed).
• Anthropic's own guidance: start with Opus 5.5 as the daily driver; escalate to Fable for ambitious open-ended work; Fable can act as adviser or orchestrator for smaller models. Theo strongly agrees on Opus 5.5.
• 264-page system card; ZDR still not available for enterprise, but a new "frontier safeguard" beta runs verification in the customer's own infra.

Benchmarks

• State of the art on Artificial Analysis (~5 points ahead of Gemini 4 Carbon), Frontier Code 2, agentic science research (74.6, up from 52.6 on 5.1 and beyond GPT-6 Astra), cyber (Mythos, ~90% CVE rediscovery), protein design, CAD.
• Gains concentrate in the sciences beyond computer science.
• Don't use Max reasoning for code: Fable 5.2 Max scores worse than XHigh on most coding tasks while costing more. Fable High ≈ Opus XHigh in price; Fable XHigh ≈ Opus Max.

Theo's own evals

• Codebase analysis (TS-Rust, T3 Code, Lakebed): Fable 5.2 underperforms Opus and Sonnet, though it beats OpenAI models overall. On TS-Rust it made 4 wrong claims out of 61 (~$23); GPT-6.1 Soul made 38 claims, none wrong, for $4.19 in 29 min.
• Code review: good and less nitpicky — if Fable approves a PR it's probably fine — but very slow (~120 min for 8 PRs vs 20-30 min for GPT models).
• "Wise owl problem": tries to be clever, like the smart colleague who half-listens in a meeting — usually sensible, occasionally nonsense, sometimes genius.

Hands-on experience

• Needs more human-in-the-loop than Opus: it doesn't get blocked, but it spirals on unimportant decisions in ways obvious to a human.
• Good: blunt, correct answers where smaller models ramble (identified a VPS vs local-NVMe slowdown in one sentence where Terra produced slop); pushes back on bad advice from other agents and corrects itself when given sources.
• Bad: a redesign that failed to visualise the requested comparison; a "safe mode" dashboard banner that leaked the email it was meant to hide; an explainer page without the requested architecture diagram.
• Frontend: decent with Anthropic's design skill; without it, generic "Claude slop" templates.
• Still emits unnecessary eyebrow text / subheadings; fixable via AGENTS.md. Better prose and copywriting.
• Fish-slop game test: better graphics and coherence than the Opus version, but Opus made the more fun game. The Opus + Fable combo works best.
• A friend's framing: much more capable, not much more autonomous — no better taste or judgement on human decisions.

Subscription math ($200 plan)

• Separate 7-day Fable limit; 2 points of Fable usage = 1 point of the weekly limit, so maxing Fable leaves 50% of the week for other models.
• Launch reset (1-3 depending on account age) restores the Fable limit and half the weekly limit without moving the reset timer, unlike Codex resets.
• Opus 5.5 only: ~$14.5-15k of API-equivalent usage per month (~73x). Fable then Opus: ~$3k Fable + ~$7k Opus, ~$4k less in total; Fable-only multiplier drops to ~15x. Turns lost are only ~20% because Fable is cheaper and finishes threads more often.
• The API prices are inflated; enterprise margins subsidise subscriptions.

Adoption and speculation

• Opus 5.5 took over T3 Code usage: 63%+ of turns, now ~78% Claude overall (Slopolytics). ZDR compatibility drove enterprise adoption.
• Fable 5.2 runs at ~106 TPS vs Opus 112 (5.1 was 81-84), so Theo is convinced it is a smaller or cheaper-to-serve model than 5.1; the big internal "model 4" is far beyond it.
• Why 5.2 and not 5.5: Opus 4.5 and 5.5 each became his everything-model; Fable 5.2 didn't. The Opus→Fable gap is smaller than Sonnet→Opus, so the version number signals that it isn't the big jump.
• Conclusion: Anthropic holds the best big model and the best workhorse. Fable is the brilliant but awkward expert; Opus is the bodybuilder that moves anything.