Claude Fable 5 Release Deep-Dive — Features, Benchmarks, Pricing, the Mythos Difference, and a New Safety Design

On June 9, 2026, Anthropic released a new model, "Claude Fable 5." It unleashed, for the first time in a form ordinary users and developers can use, capability at the level of "Mythos" — the frontier model long considered its most powerful internally. Anthropic positions it as "the most powerful model it offers generally." The tagline: "built for long-running, complex work." Indeed, there are reports of 12-hour autonomous runs and a multi-million-line code migration finished in a single day.

This article digs into this important release thoroughly, yet in a way beginners can follow. What Fable 5 is, how it differs from its twin Mythos 5, the benchmark numbers, pricing and availability, and the highlight of this release — its new safety design. Note that the figures and specs are quotations of Anthropic's announcement and various reports (as of June 2026); availability and pricing may change. See also the prior-generation Claude Opus 4.8.

*This article reflects June 2026. The figures, specs, availability, and pricing are quotations of Anthropic's official announcement and various reports, and may change. Always check the latest official information before use.

1. What is Claude Fable 5? — in one line

The simplest possible answer first. Claude Fable 5 is a new AI model Anthropic touts as "the most powerful it offers generally." It is built to go all-in on one thing in particular: "thinking for itself and seeing a long, complex job through to the end."

The point is that Fable 5 is optimized for "finishing a marathon," not "the cleverness of a single Q&A." Anthropic explains that "the longer and more complex the task, the larger Fable 5's advantage." In other words, this is a model designed to be the "main engine" of the AI-agent era. The difference stands out most when you hand autonomous work to an AI agent.

2. Fable vs. Mythos (same brain, different safeguards)

This release is two models at once — "Fable 5" and "Mythos 5." It is easy to get confused, so let us make the relationship clear. The two are actually exactly the same model inside. What differs is only the safeguards.

Anthropic itself states that "the safeguards are what distinguish the two (and why they have different names)." The same powerful brain gets a different name depending on who it is opened to, and how far — this design is the most noteworthy point of the release (more in section 7). Just remember that what we, ordinary users, touch is Fable 5.

3. The benchmarks

So how smart is it? Look at the representative benchmarks (performance tests) Anthropic published. It stands out especially on software engineering and long-running analytical tasks.

*All are quotations of Anthropic's announcement and various reports (as of June 2026). Benchmarks vary with measurement conditions and do not always match real-use experience.

Note the 80.3% on SWE-Bench Pro. This measures how well a model solves hard tasks close to real software development; it jumped from the prior-gen Opus 4.8 (69.2%) and pulled clearly ahead of GPT-5.5 (58.6%). On some benchmarks it beats Opus 4.8 by more than 10%. With a new state-of-the-art in vision (image understanding) as well, it leads on "code, text, images, and long-running tasks" alike — a flagship in the literal sense.

4. Its real strength: long-running, autonomous tasks

More striking than the benchmark numbers is its "power to keep running for a long time." Fable 5 holds its focus across millions of tokens (an enormous context) and clears, in one go, work that would take humans months.

This "marathon strength" ties directly to the era of handing work to AI agents. Beyond answering short prompts, it keeps cycling "research → plan → implement → verify" by itself for hours — and its completion rate and stability are Fable 5's real selling point. As a mechanism for running long tasks, it also pairs well with autonomous loops like Claude Code's /loop command. On the other hand, precisely because of that powerful autonomy, attention to agent security (permissions, misoperation) matters more than ever.

5. Pricing, context, availability

Let us sort out "how much, and how to use it." The pricing feels held in place given the performance (reported as half the Mythos preview), and the context window is large.

The in-app banner "Included in your plan limits until Jun 22" is exactly this notice that "you can try it within your plan, at no extra cost, until June 22." In other words, it is a "trial period" right now. For Pro/Max and similar subscribers, it is a perfect time to touch the strongest model without worrying about extra fees. From the 23rd, it is expected to consume usage credits. For a price comparison across models, see the Claude pricing comparison.

6. Compared with Opus 4.8 (price & performance)

With Fable 5 here, a new question many people have is "so which should I use — this, or the prior-gen Opus 4.8?" Let us line up the current (June 2026) pricing and performance head-to-head.

*Quotations of Anthropic's official info and various reports as of June 2026. Opus 4.8 also has a ~2.5× faster "Fast Mode" at $10/$50 per 1M tokens — the same as Fable 5. Prices and specs may change.

The takeaway is simple. Fable 5 costs about 2× Opus 4.8 (standard), but on a hard coding metric it leads by +11.1 points (80.3% vs 69.2%), and the gap widens the longer and more complex the task. The context window (1M tokens) and max output are the same. A rough guide to choosing:

What is interesting is that Opus 4.8 is also the "safety net" Fable 5 falls back to in dangerous areas (next section). So the two are rivals and, at the same time, complementary in role. Rather than "always use the strongest," the cost-and-performance optimum is to split smartly — Fable 5 for heavy, complex work, Opus 4.8 (standard) for the everyday. See the Claude pricing comparison as well.

7. A new safety design (3 classifiers)

This is the biggest highlight of the release. Why could an "in-house top-tier" model be released publicly? The answer is a new safety design that pinpoint-stops only the dangerous areas. Fable 5 has three classifiers (checkers) built in.

The clever part is that it stops "only when it is dangerous." According to Anthropic, these safeguards trigger in under 5% of sessions. That means more than 95% of sessions get Fable 5's full performance as-is. "Strongest as usual, with the brake only at the dangerous moment" — this pinpoint control is the core of what made a public release of such a powerful model possible. A new policy also retains all Mythos-class traffic for 30 days for safety purposes.

8. Days after warning "AI is too dangerous"

There is a context here you should not miss. Anthropic is reported to have released Fable 5 just days after warning that "AI is becoming dangerous." This seemingly contradictory move is exactly what symbolizes the company's strategy.

Right or wrong, this "open and close capability by area" approach may become one template for future frontier-model releases. As AI grows ever stronger, the historic significance of Fable 5 may lie in showing a third path — not the binary of "open everything / close everything." If the debate about AGI and beyond interests you, see What is AGI? too.

9. When to use it, and how to start

Finally, the practical side. It is not the case that "you should always use Fable 5." It is powerful but also priced at the top. Splitting by task is the smart move.

Starting is easy. Just switch from /model in the Claude app or in-app (as the opening banner says). Developers can specify claude-fable-5 via the API, or pick it in GitHub Copilot. Since it is included with no extra charge on many plans until June 22, the best move is to throw one of your heavy tasks at it — a long investigation or a tedious coding job. The more you test it on "a long job" rather than "a short question," the better you will see Fable 5's true value.

Summary

Here are the key points of Claude Fable 5, condensed.

• What it is: Anthropic's most powerful generally available model, putting in-house top-tier "Mythos" capability into public hands safely. Built for long, complex, autonomous tasks.
• Fable vs. Mythos: Identical inside; only the safeguards differ. The public uses Fable 5.
• Performance: SWE-Bench Pro 80.3% (Opus 4.8: 69.2 / GPT-5.5: 58.6), a first-ever 90%+ on Hex, and new SOTA in vision.
• Real strength: Focus across millions of tokens, 12-hour autonomy. Stripe completed a 50M-line migration in one day.
• Pricing/availability: $10/$50 per 1M tokens, 1M context. Free on each plan until June 22, via API and Copilot.
• Vs. Opus 4.8: About 2× the price, +11.1 points on SWE-Bench Pro; split — Fable 5 for heavy work, Opus 4.8 (standard) for the everyday.
• Safety: Cyber / bio-chem / distillation classifiers fall back only when dangerous; 95%+ run on their own. One answer to "AI is dangerous."

In the end, the essence of Fable 5 is "an AI that runs longer and finishes on its own," not just "a smarter AI" — and at the same time, a new design answer to "how to reconcile power and safety." It clearly advances the shift of AI from "a tool that answers" to "a partner you delegate to." If you are a Pro/Max user, throw your most tedious job at it while it is free to try. For the comparison with the prior generation, start from the Claude Opus 4.8 explainer.

FAQ