Claude Fable 5, explained: what Anthropic actually shipped (and the caveats)

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.

What did Anthropic actually ship on June 9?

On June 9, 2026, Anthropic shipped Claude Fable 5 — generally available, model ID claude-fable-5 — alongside the restricted Claude Mythos 5. Fable is the one you can call today. If you build on Claude in production like I do, claude-fable-5 is the string you’ll be typing into config files for the next while.

Two models landed that day, and the difference matters. Fable 5 is the one you and I can actually use — generally available, safeguarded. The other, Claude Mythos 5, is not general. Anthropic limited it to “Project Glasswing” approved partners plus a handful of select biology researchers, and for those authorized users it ships with lifted cybersecurity guardrails. So when you read the big launch numbers, keep asking: is this Fable, the model I can call, or Mythos, the one behind the velvet rope?

My framing: Fable 5 is the safe-for-everyone cut of Mythos. Same family, capped for the public. That single idea explains almost everything quirky about how it behaves, including the fallback I’ll get to below.

What is the new “Mythos-class” tier above Opus?

Here’s the part that’s easy to miss in a version-number world. Anthropic says Mythos-class models “sit above our Opus class in capability.” That’s a new top tier, not Opus 4.9 with a fresh coat of paint.

So the mental model is: Mythos is the new ceiling, and Fable 5 is the safeguarded version of Mythos made safe for general use. You’re getting access to a higher tier than Opus — with guardrails bolted on.

For builders, the practical takeaway is about model selection. Your “use the best Claude” default just moved up a rung, and the price moved with it. Fable isn’t a drop-in Opus swap where you flip a flag and pocket free intelligence — it’s a deliberate “I need the top tier and I’ll pay for it” choice. I’ll come back to when that math works.

How good is it really? The benchmarks

It launched #1 on the Artificial Analysis Intelligence Index. That’s the headline, and it’s earned — but let me give you the numbers that actually move my decisions.

On SWE-Bench Pro, which is the one I watch for agentic coding, Fable 5 hit 80.3%. GPT-5.5 sits at 58.6% — that’s a 21.7-point gap — and Fable is roughly 11 points above 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%, with Gemini at 0.0%). FrontierCode is roughly 2x Opus 4.8. Hebbia’s finance benchmark came back as the highest score of any model. And vision is now genuinely state-of-the-art — it can play Pokemon FireRed from vision alone.

The honest caveat, and Anthropic’s reporting is upfront about this: some of the top benchmark rows reflect Mythos 5, not Fable 5. So when a chart shows a frontier-leading number, check which model it’s measuring. The Fable 5 you and I can call is excellent — but it’s not always the row sitting at the very top.

How much does it cost, and where does it run?

Fable 5 runs $10 per million input tokens and $50 per million output tokens. That’s about 2x Opus 4.8, which is $5/$25. The 90% prompt-caching discount brings cached input down to roughly $1 per million — and Anthropic notes Fable is “less than half the price of Claude Mythos Preview.” So it’s premium, but it’s the cheap end of the new tier.

The context window is 1,000,000 tokens. Inputs are text, image, and file; output is text, with reasoning and extended thinking supported. It runs on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, Microsoft Foundry, and claude.ai.

One scheduling detail worth flagging because it bit people this week: on claude.ai subscription plans (Pro, Max, Team, and seat-based Enterprise), Fable 5 was free — no extra cost — through June 22, 2026. Starting June 23, 2026, it requires usage credits “until capacity allows standard inclusion.” On the API and consumption-based Enterprise, it’s fully pay-as-you-go from the start. If your free Fable suddenly went quiet this week, that’s why. For how token pricing translates into a real monthly bill, I’ve broken that down in how much it costs to add AI to your app.

How does the Opus 4.8 safety fallback work?

This is the most unusual thing about Fable 5, and you need to understand it before you architect anything on top of it.

Classifiers watch for three topic areas — cybersecurity, biology/chemistry, and model-distillation. When one trips, the response is, in Anthropic’s words, “automatically handled by Claude Opus 4.8 instead,” and the user is notified. So Fable doesn’t refuse on those topics; it quietly hands the turn to a different, lower-tier model.

Anthropic says more than 95% of Fable sessions involve no fallback, so for most work you’ll never see it. On the safety record: an external bug bounty found “no universal jailbreaks in over 1,000 hours of testing,” and Anthropic acknowledged “there could still be novel attacks.”

Here’s my builder caveat, and it’s a cost one. On flagged topics, Fable 5’s real-world performance is effectively Opus 4.8 — because it literally is Opus 4.8 answering. But you’re still on the Fable tier. So if your workload leans into security tooling or anything bio-adjacent, you can end up paying roughly 2x for Opus-quality output. Know your domain before you commit. (If you’re new to where any of this sits, what is an AI agent is the primer I point people to.)

Can it really run autonomously for that long?

This is the capability I actually care about. Anthropic says Fable 5 “can work autonomously for longer than any previous Claude models” and “stays focused across millions of tokens in long-running tasks.”

The proof point that made me sit up: Stripe used Fable 5 to do a codebase-wide migration on a 50-million-line Ruby codebase in one day — work Anthropic says would have taken a team over two months by hand. Cognition (via FrontierCode) and Hebbia (finance) also put it through their paces.

My take: long-horizon focus is the feature that changes how I’d architect agents, more than any single benchmark. When a model holds its thread across millions of tokens, you stop building elaborate scaffolding to chop a job into tiny supervised steps and re-stitch the results. You hand it the whole task. That’s a different shape of system — and it’s the part of this release I’m most interested in building on. More of how I think about that is in my AI engineering guides.

The controversy: caution vs commercialization

I said I wouldn’t write a fanboy post, so here’s the uncomfortable part.

Anthropic released Fable 5 to the public just days after warning that AI is becoming too dangerous. The company had urged major labs to adopt “a coordinated brake pedal on frontier AI development,” cautioning that systems may soon reach recursive self-improvement — improving themselves without human oversight. Then it shipped a model above its previous ceiling to everyone. Critics frame that as a contradiction between stated caution and aggressive commercialization, and I don’t think that read is unfair.

Then there’s the retention policy. A mandatory 30-day data-retention policy applies to all Fable and Mythos traffic — even customers with zero-retention agreements — framed as necessary to defend against “novel attacks.” Reporting notes this could set a precedent: access to powerful models bundled with mandatory retention, justified as “safety.” If you operate under data agreements that promised zero retention, this is a contract and compliance question, not a footnote. Read it before you route production traffic through Fable.

FAQ: quick answers for builders

What’s the model ID? claude-fable-5.

Is Mythos 5 available to me? No. It’s limited to Project Glasswing approved partners and select biology researchers.

What’s the context window? 1,000,000 tokens. Inputs are text, image, and file; output is text, with reasoning/extended thinking.

Why did my flagged request feel like Opus? Because it was. Cybersecurity, biology/chemistry, and distillation topics are auto-handled by Opus 4.8, and you’re notified when it happens.

Is it still free on my plan? It was free on subscription plans through June 22, 2026. From June 23, it needs usage credits until capacity allows standard inclusion.

My bottom line

The state-of-the-art is real. #1 on the Artificial Analysis Intelligence Index, 80.3% on SWE-Bench Pro, and long-horizon autonomy I can genuinely build on — that Stripe migration isn’t a demo, it’s the workload I want to point this at.

But the trade-offs are just as real. You’re paying about 2x Opus, and on flagged topics you’re paying that 2x for Opus-quality output, because the fallback hands the turn to Opus 4.8. The caution-versus-shipping tension and the mandatory 30-day retention are things to watch, not wave away — especially if you carry zero-retention commitments.

Honest verdict: I’d reach for Fable 5 on hard, long-running agentic coding work where the autonomy earns its keep. I’d think twice for flagged domains, where I’m paying a premium for a model I’m not actually getting, or for cost-sensitive jobs where Opus 4.8 does the work for half the price. If you’re weighing Fable for a production build and want a second set of eyes on the cost-versus-capability math, that’s the kind of work I do.