July 25 update: Claude Opus 5 now ranks above Fable 5 here. Fable still keeps narrow peak wins on FrontierCode, CursorBench, no-tools HLE, and a held-out legal test, but Opus 5 leads more broadly relevant agent and knowledge-work evaluations at half the token price and with fewer deployment restrictions. The review below preserves Fable’s June launch evidence; read its superlatives in that historical context.
If Opus 4.8 was the promotion, Fable 5 is the corner office. Anthropic’s naming shift from musical tiers (Haiku, Sonnet, Opus) to literary ones (Fable, Mythos) isn’t just branding — it signals a new class of model. Fable 5 runs on the same Mythos-class architecture that powers the restricted Mythos 5, but with safety classifiers that make it safe for general use. Think of it as a supercar with the speed limiter set — still the fastest thing on the road, just with guardrails on certain turns.
The numbers tell the story. SWE-Bench Pro 80.3% doesn’t just beat GPT-5.5 (58.6%) — it embarrasses the entire field. FrontierCode Diamond at 29.3% means Fable 5 writes production-quality code five times more efficiently than GPT-5.5 (5.7%). On Hebbia’s Finance Benchmark — senior-level document reasoning, chart reading, root-cause analysis — it’s #1. On CursorBench, it opened up “a class of long-horizon problems that were out of reach for earlier models.”
But the most telling demonstrations aren’t benchmarks. Stripe migrated a 50-million-line Ruby codebase in one day — work that would have taken a full team two months. The model completed Pokémon FireRed using only raw screenshots — no maps, no helper tools, no game-state data. And when given persistent file-based memory playing Slay the Spire, its performance improved 3× more than Opus 4.8’s.
The safety and deployment story is worth understanding. Following temporary export control reviews in mid-June, Fable 5 was fully restored globally on June 30, 2026. Shortly after, Amazon security engineers credited Fable 5 with autonomously identifying and patching a zero-day authentication vulnerability across 12,000 internal services in hours. To prevent dual-use misuse, queries touching advanced cybersecurity exploit generation, biology, chemistry, or model distillation get automatically routed to Opus 4.8 via a specialized classifier. This happens in less than 5% of sessions, though it can introduce false positives on benign systems programming. The unrestricted Mythos 5 is reserved for vetted partners through Project Glasswing — where it’s already helping defend critical software infrastructure.
The real question is whether the price is worth it. At $10/$50 per million tokens, Fable 5 costs roughly 2× what Opus 4.8 does. But token efficiency partially offsets this — achieving FrontierCode-leading results at medium effort means less compute per task. For professionals whose time is worth more than their API bill, the math is simple. For everyone else, Opus 4.8 remains excellent. But if you want the best generally available AI model on the planet — the one where the gap widens as the task gets harder — this is it.
