Fable 5: Mythos on a Leash

Anthropic spent the week telling the industry to find a brake pedal. They published an essay urging the labs to coordinate a slowdown on frontier AI, warning out loud about recursive self-improvement — models that get better at building models without us in the loop. Then, days later, on June 9, they shipped the most powerful model Anthropic has ever handed the public.

That model is Fable 5. And the way they squared the circle is the whole story: Fable 5 is a Mythos-class model — a tier above Opus — wearing a leash. Understand the leash and you understand everything about how to use it, what it costs you, and why Anthropic built it this way.

What Fable 5 actually is

For most of the last year the top of Anthropic's stack was Opus. Fable 5 sits above it. It is the first generally available model from the Mythos family — the internal line that spooked governments and the cybersecurity world earlier this year for superhuman vulnerability-finding, and that Anthropic has been supplying to vetted critical-infrastructure partners across fifteen countries under tight control.

Here is the part people keep missing: Fable 5 and Mythos 5 are the same model. Same underlying model, same capability class. The only difference is that Mythos 5 has the safety classifiers lifted for vetted access, and it is not generally available — it ships in limited release through a program called Project Glasswing to vetted, mostly government and critical-infrastructure customers. So when you call claude-fable-5, you are talking to Mythos with a muzzle on. Fable is the gated public face of a model Anthropic does not fully trust the open internet with.

That framing matters because it tells you the capability ceiling is real, not marketing. The benchmark gap to Opus 4.8 is not a tuning artifact — it is a bigger, more expensive model that Anthropic decided to expose carefully rather than freely.

The specs that matter

A 1-million-token context window by default — no beta header, no opt-in dance. Up to 128K output tokens per request. Knowledge cutoff of January 2026. And pricing that will reset your cost model: $10 per million input tokens and $50 per million output — exactly double Opus 4.8. Cached reads land around $1 per million, the usual ~90% prompt-caching discount, which on a model this expensive stops being a nice-to-have and becomes load-bearing.

The leash is the architecture

Fable 5 ships with classifiers watching three categories: cybersecurity (exploitation, vulnerability discovery, agentic hacking), biology and chemistry (dual-use bioweapon work, certain gene-therapy design), and — the interesting one — distillation: attempts to pull the model's capabilities out to train a competitor. That third category is the quiet strategic tell. Anthropic isn't only guarding against misuse; it's guarding against its own capability leaking out as someone else's training data. Flag any of those and the model declines.

What happens next is the cleanest part of the design. A refusal is not an error. The Messages API returns it as a successful HTTP 200 with stop_reason: "refusal", and it tells you which classifier declined. From there you can fall back to Opus 4.8 — still under the standard Claude safety stack, just without Fable's Mythos-class classifiers. There is a new fallbacks parameter (beta) that retries server-side, and SDK middleware that does it client-side across TypeScript, Python, Go, Java, and C#. You are not billed for a refusal that produced no output, and a “fallback credit” refunds the prompt-cache cost of switching models so you do not pay the cache tax twice.

Anthropic says fewer than 5% of sessions ever trip a classifier; 95%+ run on full Fable. So for ordinary engineering work the leash is invisible. But notice what they actually built: not a model that refuses, but a model that refuses gracefully and hands you a cheaper, weaker model to finish the job. The fallback is not an edge case to handle — it is part of the product. You are meant to wire Opus 4.8 in as the catch.

The price of the leash: you give up zero-data-retention

This is the cost nobody puts on the pricing page. Fable 5 and Mythos 5 are designated Covered Models, which means 30-day data retention and no zero-data-retention option. Opus, Sonnet, and Haiku can run ZDR; Fable cannot. The classifiers need to retain prompts and outputs to do their job, so the safety mechanism and the retention policy are the same decision. Access is logged, data is deleted after 30 days barring safety investigations or legal holds, and Anthropic says it is not used for training — but if your compliance posture depends on zero retention, Fable is off the table, full stop. GitHub Copilot makes this explicit: the Fable policy is off by default for orgs precisely because it breaks the ZDR they grant the other Claude models.

What changes in the API (read this before you swap the model string)

If you are migrating from Opus 4.8, the request surface has a few hard edges. Adaptive thinking is the only thinking mode — thinking: {type: "disabled"} returns a 400; just omit the parameter to run without it. Sampling parameters (temperature, top_p, top_k) are gone, same as 4.7/4.8. The raw chain of thought is never returned — thinking.display defaults to "omitted"; set "summarized" if you want readable reasoning. Depth is controlled by the effort parameter (low / medium / high / xhigh / max), and Task Budgets, the memory tool, context editing, compaction, and vision are all supported. Pass thinking blocks back unchanged in multi-turn calls on the same model.

None of this is exotic if you already run Opus 4.7 or 4.8 — it is the same surface. The genuinely new things to handle are the refusal stop reason and the retention change, not the thinking config.

Is it actually better?

On the numbers, yes, and not subtly on the work it is built for. SWE-Bench Pro lands at 80.0% (Mythos 5, unmuzzled, edges it to 80.3) — for context, GPT-5.5 sits at 58.6%, and Fable is roughly eleven points over Opus 4.8's 69.2. FrontierCode Diamond, Cognition's hard agentic-coding eval, is 29.3% versus Opus 4.8's 13.4% — better than double. Finance reasoning and vision are described as state of the art, with the biggest gains showing up on long, complex, autonomous tasks rather than one-shot questions.

The independent reception lines up with that shape. Simon Willison spent five and a half hours with it, called it “a beast,” and concluded the real challenge was finding tasks it couldn't do — he had it convert a MicroPython library to full CPython-on-WebAssembly and refactor a tool-approval system into clean library features, work he estimated would have taken him days. He also thinks it is a very large model, “maybe the largest yet from any vendor.” The catch came on the invoice: $110 in a single day, most of it one agent session. The other half of the internet landed on a sharper line — “a Ferrari with a 30mph limiter” — aimed squarely at the safeguards.

The honest read from people who tested it: ordinary conversation does not feel dramatically different from Opus 4.8. Hard frontend and long-horizon agentic coding clearly do. The gap is real and reproducible exactly where the price premium hurts least relative to the value — and invisible everywhere else, where you are just paying double for a vibe.

The billing play

Watch the calendar, because the packaging is doing work. From June 9 through June 22, Fable 5 is included free on Pro, Max, Team, and seat-based Enterprise — you just pick it in the model picker, no extra charge. While it sits in your plan it counts roughly double an Opus session against your limits, so it burns your monthly allowance about twice as fast.

On June 23, Fable is set to come off the subscription plans — and here is exactly what that means for your bill, because the phrase “usage credits” hides the part you actually want to know. Your Max plan is never charged automatically. To use Fable after the 22nd you have to opt in: in Settings → Usage you switch on extra-usage credits. Leave it off and Fable simply isn’t available to you — you stay on your flat Max fee with the plan’s included models (Opus, Sonnet, Haiku) and nothing extra touches your card. Switch it on and every Fable token draws down credits at full API rates ($10 per million input, $50 per million output), billed to your payment method as a separate charge on top of Max, with a monthly spend cap you set (or “unlimited”) and an optional auto-reload that is also off by default. So “Max plus Fable after June 22” really means “your Max plan plus a pay-per-token meter you switched on yourself.”

Read it for what it is: a two-week taste, then a deliberate nudge toward usage-based billing. That is not a knock — capacity on a model this large is genuinely scarce and metering it is rational — but if you wire Fable into a workflow this month, price the workflow as if the free window is already closed. Heavy agentic sessions routinely run 500K to a million tokens, and at $50 per million output that is real money per task, not rounding error. Anthropic says it intends to fold Fable back into standard plans once capacity allows, but has given no date.

How to actually turn it on

In Claude Code: run claude update to get v2.1.170 or later, then /model fable (or /model claude-fable-5). It will not appear in the picker on older versions. In the API and on the Console, the model id is claude-fable-5 — it is in the model dropdown, and the 1M context window is the default, so there is no header to set. On claude.ai it is in the model picker on Pro/Max/Team (free until June 22). It is also live on Claude Platform on AWS, Amazon Bedrock, Vertex AI, Microsoft Foundry, and GitHub Copilot (Pro+/Max/Business/Enterprise, with the org policy off by default).

On the “Fable 5” versus “Fable 5 (1M context)” question: for this model the million-token window is the default, so you get the full context either way. On the API there is nothing to set; in Claude Code, /model fable gives you the 1M window. The [1m] suffix you see on some models is a Claude Code convenience for pinning the long-context variant explicitly — on Fable it is redundant. If you ever want to strip the 1M variants out of your Claude Code picker, CLAUDE_CODE_DISABLE_1M_CONTEXT=1 does it.

The operator's read

Do not marry Fable 5. Route to it. The correct mental model is a two-model system you wire on purpose: Fable for the long-horizon, high-stakes, agentic work where the capability gap pays for the price premium, and Opus 4.8 as the default workhorse for everything else — cheaper, ZDR-capable, and, conveniently, the exact model the fallback already hands you. The classifier-plus-fallback design is basically Anthropic shipping that routing decision for you and charging you to opt out of the cheap side.

Here are the takes worth arguing about:

1. The headline isn't the benchmark bump — it's that Anthropic shipped a model it doesn't fully trust the public with, and made the leash a first-class API feature. That is a new product category: capability you rent through a safety gate.

2. Giving up zero-data-retention is the real price, and it is buried. For regulated teams that is a harder no than the dollar cost.

3. “Warn the industry to slow down, then ship the most powerful public model with a fallback leash” is not hypocrisy — it is the actual thesis. The leash is how they reconcile the two. Whether you buy it is the interesting question.

4. The free-then-credits window is a funnel. Treat the free fortnight as a demo, not a budget.

5. Most teams should keep Opus 4.8 as the default and call Fable for the 5% of tasks that actually need a tier-above-Opus model. Reaching for Fable on a summarization job is paying frontier prices to sit in traffic.

6. Mythos 5 existing at all — same weights, no leash, government-only — tells you the public ceiling and the real ceiling are now different models with the same name. That is the line everyone will be staring at for the next year.

7. The next operator skill isn't prompting — it's traffic control. Which model, which retention policy, which effort level, which fallback, which cache prefix, which task budget. Fable doesn't just add a faster engine; it makes the routing layer the job. Read the benchmark scorecard as what it actually is: a routing table.

Fable 5 is the most capable model Anthropic lets you call from a terminal today. It is also the clearest signal yet that the frontier is being handed out in measured doses, behind classifiers, with the bill metered by the token. Use it for what it's for, keep Opus 4.8 wired as the default, and remember what's actually on the other end of that API call.