Розбір

Claude Fable 5: strong, expensive, with an asterisk

Anthropic shipped its most capable public model — right after warning that AI is too dangerous. A no-hype breakdown: what's real, what's marketing, two integration traps, and how to test it yourself.

Oleksandr Polianskyi

— 4 хв читання4 min read
Claude Fable 5: strong, expensive, with an asterisk

On June 9–10 Anthropic shipped Claude Fable 5 — its most capable publicly available model. A few days earlier the same company warned that AI is becoming too dangerous. That contrast alone is worth a pause. Let's break it down without the hype — and, more importantly, what it means once you actually integrate it into a product.

The short version

  • Fable 5 (claude-fable-5) — Anthropic's most capable public model. Alongside it, Mythos 5 (claude-mythos-5) — the same power without the safeguards, for select customers only (Project Glasswing).
  • 1M context, up to 128k output. $10/$50 per 1M (output ~twice as expensive as Opus 4.8).
  • Adaptive thinking is always on; the raw chain-of-thought is never returned.
  • Free on Pro/Max/Team June 9–22.

"SOTA on nearly every benchmark" — whose number is that?

The model is genuinely strong — Simon Willison confirms it independently. But "nearly every" means not every (the ones it loses, they don't show), and a chunk of the benchmarks are internal or customer-run (Hex, IMC, Hebbia): the seller, or the buyer, measuring itself. Independent reproductions on launch day are scarce.

"Stripe migrated 50 million lines of Ruby in a day"

The loudest number — and the most manipulative. In reality it's a single codebase-wide migration (one kind of mass transformation) in a 50M-line base, not a language rewrite. It's the class of task where deterministic codemod/AST tools already worked before LLMs. And it's "Stripe reported" inside Anthropic's press release — without an engineering post from Stripe itself. Who reviewed the result, what it cost — silence.

You pay twice: money and predictability

$10/$50, slower than Opus, easy to burn over $100/day — Willison burned $110.42 in one intensive day, more than a monthly subscription. Add mandatory 30-day retention (no zero-data-retention) — GitHub made Fable off-by-default for Business/Enterprise because of it. For proprietary code that's a compliance factor you didn't have with Opus.

Trap #1: refusals are normal behavior, not an error

Classifiers decline requests in three domains (cybersecurity, biology/chemistry, model distillation) and serve the answer from the weaker Opus 4.8 instead. The key part: a refusal arrives as stop_reason: "refusal" — an HTTP 200, not an error. If your code only checks the HTTP status, it accepts a "successful" response that is actually a refusal.

const res = await client.messages.create({ model: "claude-fable-5", ... });
if (res.stop_reason === "refusal") {
  // not an error — the model declined. Retry on another model:
  return await client.messages.create({ model: "claude-opus-4-8", ... });
}

Anthropic gives three ways to fall back: a server-side fallbacks parameter, SDK middleware, or manual. You aren't billed for a refusal before generation, and fallback credit refunds the prompt-cache cost on the switch.

Trap #2: invisible degradation

Separate from the visible redirect: ~0.03% of queries around frontier-AI development quietly lose quality — no refusal, no signal, no fallback. In a paid product that means an unpleasant thing: you can't always tell the model's own limit from a hidden vendor intervention. This isn't a conspiracy theory — it's documented behavior; it just doesn't make it into the loud announcements.

For security engineers — separately, and painfully

"Approved" status does NOT lift the offensive filter. Legitimate pentesting, vuln research, malware analysis for defense — all of it trips the classifier and quietly downgrades you to the weaker model. The Register caught refusals on a bare "Hello"; Anthropic admitted the over-tuning and promised to reduce false positives. And jailbreaks are already public (Pliny bypassed the classifiers and even extracted the system prompt). So the barrier stops the lazy and penalizes the honest — an asymmetry that doesn't favor the defender.

The "AI is too dangerous" narrative right before launch

Not necessarily hypocrisy. Sincere concern and commercial interest don't contradict here — they coincide: "it's dangerous, but we're the only ones who can do it safely" = the justification for a premium price and an exclusive Mythos for the select few. Safety becomes both a product and a competitive moat. Skepticism is warranted not because someone is lying, but because the incentives are aligned exactly that way.

How to check for yourself (measure, don't trust benchmarks)

  • Cost/latency: run your typical agentic task 20 times, measure $/task and p95 latency against Opus 4.8.
  • False-positive audit: a batch of 100 legitimate security prompts (reading a CVE, analyzing malware for defense, a fail2ban config) → count the % of silent or visible downgrades.
  • The migration myth: take your own codemod task (typing, an API rename across the repo) and compare Fable 5 vs a deterministic AST tool on correctness and cost — "no AI" often wins.

When it's actually worth it

  • Hard tasks where quality gives a measurable win and $50/1M pays for itself in the result.
  • Large contexts (1M) where a simpler model drowns.

When it's NOT

  • Routine, where Opus 4.8 already handles it.
  • Security/offensive work — the filter will break your process.
  • Anything where predictable cost and behavior are critical.

The practical minimum before integrating: handle stop_reason: "refusal" + fallback, alert on daily spend, and don't drop Opus 4.8 from the stack — here it's both your fallback and the silent substitute under the hood.

Sources

This is my experience and view, not dogma. The model is fresh and independent benchmark reproductions are still scarce — verify what's current for your own task. Facts as of 2026-06-11.

БільшеRead more