Claude Fable 5: The Best Model Anthropic Told Us Not to Build
Fable 5 is the most capable public Claude ever — it tops SWE-Bench Pro at 80.3%. But Anthropic shipped it days after telling the industry to hit the brakes, attached mandatory 30-day retention to every request (zero-retention agreements included), and quietly routes flagged topics to Opus 4.8 at Fable prices. Use it for big long-horizon jobs. Route everything else to cheaper models.
Update — June 30, 2026: Claude Fable 5 (
claude-fable-5) is back. After the June 12 US export-control directive that had taken both models offline was lifted on June 30, Anthropic is redeploying Fable 5 globally on July 1, 2026 across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork (cloud providers to follow). It returns with new safety classifiers that block certain harmful cybersecurity tasks; on Anthropic’s own surfaces a blocked request falls back to Opus 4.8, and Anthropic flagged more false positives in routine coding and debugging that it will keep refining. See what changed in the redeployment and how to handle the classifier fallback as a developer. The analysis below describes Fable 5 as it launched; treat its availability and pricing as current again.
Anthropic shipped the model it warned us about
Read the timeline and tell me it doesn’t bother you. Days before Claude Fable 5 went live, Anthropic publicly urged the major labs to adopt a “coordinated brake pedal on frontier AI development,” warning that systems may soon reach recursive self-improvement — improving themselves without human oversight. Then, on June 9, 2026, it shipped its most powerful public model yet, generally available, model ID claude-fable-5, “available everywhere today.”
I build and ship AI features and agents for a living, so here’s my position up front, as opinion before any of the facts: Fable 5 is the strongest public Claude ever made, and the capability is not hype. But it arrives with three things that deserve scrutiny, not a shrug — the warn-then-ship contradiction, a mandatory retention policy that ignores your zero-retention agreement, and a fallback that bills you premium prices for an older model. The model is brilliant. The package around it is uncomfortable. Both are true.
Fable 5 is the safeguarded public version of a brand-new tier Anthropic calls Mythos-class, which it says sits above the Opus class. Fable 5 is the one you and I can call. Its bigger sibling, Mythos 5, is locked behind approved partners.
How good is Claude Fable 5, really?
Give the model its due, because the critique only lands if I’m honest about how strong this thing is. And it is strong.
Fable 5 launched #1 on the Artificial Analysis Intelligence Index. On SWE-Bench Pro it scored 80.3% — +21.7 points over GPT-5.5 (58.6%) and roughly 11 points above Claude Opus 4.8. Spatial reasoning is about 3x Opus 4.8 (38.6% vs 14.5%). Legal reasoning leads the field (13.3% vs GPT-5.5’s 2.1% and Gemini’s 0.0%). Vision is state-of-the-art — Anthropic says it plays Pokémon FireRed from vision alone.
Fable 5 leaves the field behind
The number that actually stops me isn’t a benchmark. Anthropic says Stripe used Fable 5 to run a codebase-wide migration on a 50-million-line Ruby codebase in a single day — work a team would have spent over two months doing by hand. That’s the long-horizon, “stays focused across millions of tokens” job that benchmarks can’t capture and that I care about as a builder. Credit where it’s due: this is real.
Two caveats before you over-read the leaderboard. First, some of the top benchmark rows are Mythos 5, not Fable 5 — check which model owns each line. Second, on cybersecurity, biology, and distillation tasks Fable 5’s safeguards make it fall back to Opus 4.8, so its real-world performance there is effectively Opus-level, whatever a topline number suggests.
Three problems, named plainly
Here’s the spine of the critique. I’ll be fair to each point, but I won’t soften it.
The warn-then-ship contradiction. Anthropic told the industry to hit the brakes, then shipped the fastest car it has ever built. I won’t call it hypocrisy outright — but you don’t get to publish a “coordinated brake pedal” warning one week and your most powerful public model the next without owning the contradiction. The gap between the caution and the commercialization is real, and critics are right to point at it.
Mandatory retention, zero-retention agreement be damned. A mandatory 30-day data-retention policy applies to all Fable and Mythos traffic — even customers who negotiated zero-retention agreements. Anthropic frames it as a defense against “novel attacks.” I get the security logic. What I won’t wave away is the precedent: access to a powerful model now comes bundled with mandatory data collection, justified as “safety.” Reporting notes this could normalize exactly that pattern — access for retention, sold as protection. That’s the part I refuse to shrug at.
Pay Fable prices, get Opus on flagged topics. Classifiers detect cybersecurity, biology/chemistry, and model-distillation prompts, and on those the response is “automatically handled by Claude Opus 4.8 instead” — you’re notified, but you still paid Fable rates. So on flagged topics, you can spend premium money and get an Opus answer. Defensible as safety design. Just know it before you budget around it.
Three problems, stated plainly
Where Anthropic earned real credit
Now the other side, because the safety effort here is not theater — and a fair critique has to say so.
More than 95% of Fable sessions involve no fallback at all, so the routing I described is the exception, not the rule. An external bug bounty found no universal jailbreaks in over 1,000 hours of testing, and Anthropic openly admitted “there could still be novel attacks” instead of overselling the result. That’s the right tone.
And the restraint where it matters most is genuine: Mythos 5 — the unsafeguarded tier, with lifted cybersecurity guardrails for authorized users — was deliberately not made general. It’s limited to “Project Glasswing” approved partners and select biology researchers. Anthropic could have shipped the strongest thing it had to everyone. It chose not to. That decision is the real safety story, and I’ll credit it without hesitation.
The price: are you paying double for the right reasons?
This is where builders actually decide, so let’s make the economics legible.
Fable 5 runs $10 per million input tokens and $50 per million output — roughly 2x Claude Opus 4.8 at $5/$25. The 90% prompt-caching discount brings cached input down to about $1 per million, the single biggest lever you have. On claude.ai and subscription plans (Pro, Max, Team, seat-based Enterprise) Fable 5 was free through June 22, 2026; starting June 23 it needs usage credits “until capacity allows standard inclusion.” On the API and consumption-based Enterprise, it’s fully pay-as-you-go.
My take, bluntly: the 2x premium only pays off when the task genuinely needs the extra horizon. Summarize tickets or route intents with this and you’re lighting money on fire. For the full picture on what frontier tokens really cost a product, I broke it down in what it costs to add AI to your app.
Token pricing: Fable 5 vs Opus 4.8
How should I decide whether to use Fable 5?
Here’s the routing logic I’d actually use — the practical payoff of this whole post.
Reach for Fable 5 on genuinely ambitious, long-horizon work: large autonomous migrations, million-token tasks where it really “stays focused,” the Stripe-shaped jobs. That’s where the premium and the autonomy earn their keep, and nowhere else.
Route everything else — and anything in cybersecurity, biology, or distillation — to cheaper models. On flagged topics you’d hit the Opus 4.8 fallback anyway, so Fable prices buy you an Opus answer. If a zero-retention agreement matters to your data policy, factor in the mandatory 30-day retention before you send sensitive traffic through it, because the agreement won’t save you. And cache aggressively — that 90% discount is how you blunt the 2x. This is the same guardrails-and-cost-caps discipline I argue for in running LLMs in production, and it’s why I keep asking whether you even need the most powerful model for the job in front of you.
Should you adopt Fable 5?
- Is the task long-horizon or autonomous?Large migrations, million-token jobs that need sustained focus → Fable is a real fit.
- Is the topic flagged (cyber/bio/distillation)?If yes, you'll hit the Opus 4.8 fallback anyway — route to a cheaper model.
- Does mandatory 30-day retention conflict with your data policy?Zero-retention agreements don't exempt you — don't send sensitive traffic if it matters.
- Can you cache to cut cost?The 90% caching discount (~$1/M cached input) is your best defense against the 2x price.
FAQ
Is Fable 5 the same as Mythos 5? No. Mythos 5 is not general — it’s limited to Project Glasswing approved partners and select biology researchers, with lifted cybersecurity guardrails for authorized users. Fable 5 is the safeguarded version made safe for general use.
What does it cost? $10 per million input tokens, $50 per million output tokens, with cached input around $1/M thanks to the 90% prompt-caching discount.
Is it still free on my plan? It was free on Pro, Max, Team, and seat-based Enterprise through June 22, 2026. Starting June 23 it needs usage credits until capacity allows standard inclusion.
What’s the context window? 1,000,000 tokens. Inputs are text, image, and file; output is text with extended thinking/reasoning support.
Where can I use it? Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and claude.ai.
What’s the retention catch? A mandatory 30-day retention policy applies to all Fable/Mythos traffic — even customers with zero-retention agreements.
My bottom line
Fable 5 is the best public model Anthropic has ever shipped — and the company spent the week before launch telling the industry to slow down. That’s the post in one sentence. The capability is real. The contradiction is real. The mandatory retention is real. And the model you pay for isn’t always the model you get.
So: worth it for the ambitious, long-horizon jobs where autonomy pays for the premium — and worth routing everything else to cheaper models. The thing I’ll watch isn’t the benchmark leaderboard; it’s the precedent. If “safety” becomes the standing reason your data gets retained, and powerful models start arriving with mandatory data collection as the price of entry, that’s the trade that outlasts any one launch. Keep your eye on it. More AI engineering guides live over in my AI agents hub.