If you've seen "Claude Fable 5 Mythos" referred to as an AI storytelling tool — including in an earlier version of this article — that description is wrong, and this page has been rewritten to correct it. Fable 5 and Mythos 5 are two separate models Anthropic released on 9 June 2026, and neither is a storytelling product. They are general-purpose frontier models built for software engineering, knowledge work, vision, and scientific research. Here's what they actually are, what they cost, and who can use them.
Correction — this article previously described a product that doesn't exist
An earlier version of this page reviewed "Claude Fable 5 Mythos" as a single storytelling platform with features like narrative memory and character-consistency tracking. No such product exists, and those features were not real. The evocative names appear to have been mistaken for a fiction-writing tool. I've replaced the article entirely rather than patching it, and I'm leaving this note up rather than quietly deleting the old version — if you read the earlier page and made a decision based on it, you deserve to know it was wrong.
At a glance
- What they are: Two frontier language models from Anthropic, built on the same underlying architecture
- The difference: Mythos 5 is the same model as Fable 5 with safeguards lifted in some areas
- Released: 9 June 2026
- Price (both): $10 per million input tokens, $50 per million output tokens
- Access: Fable 5 is generally available; Mythos 5 is restricted to vetted partners
- Not: A storytelling app, a writing assistant, or a consumer product with a monthly subscription
Two models, one architecture
The naming causes genuine confusion, so it's worth being precise. Anthropic describes Mythos 5 as "the same underlying model as Fable 5, but with the safeguards lifted in some areas." They are not a tiered product line where one is a cheaper version of the other, and they are not two halves of one product. Same capabilities, same price, different access rules.
Claude Fable 5 is the public release. Anthropic says its "capabilities exceed those of any model we've ever made generally available," with particular strength in software engineering, knowledge work, vision, and scientific research. Anyone can use it through the API.
Claude Mythos 5 is the same model with certain guardrails removed, distributed through Project Glasswing — a cybersecurity initiative Anthropic runs with roughly 50 partner organisations, including Nvidia, Amazon, Apple, Google, Microsoft, Broadcom, Cisco, CrowdStrike, JPMorgan Chase, Palo Alto Networks, and the Linux Foundation. Anthropic reports the programme has surfaced more than 10,000 high- or critical-severity vulnerabilities across widely used software since it began. You cannot sign up for direct Mythos 5 access the way you'd sign up for an API key — though as of 21 August 2026 there's a narrower second path, covered below.
That split is the most interesting thing about this release, and it's worth sitting with. Anthropic built one model, then shipped two versions of it: a restricted one for the public and an unrestricted one for vetted institutions working in exactly the domains the restrictions cover. Whatever you think of that arrangement, it's an unusually explicit statement about where the company thinks the risk sits.
What it costs
Both models are priced identically: $10 per million input tokens and $50 per million output tokens.
Anthropic notes this is less than half the price of Claude Mythos Preview, the earlier invitation-only model this line succeeds. But context matters more than the direction of travel: this is still Anthropic's most expensive model by a wide margin. For comparison, Claude Opus 5 — which succeeded Opus 4.8 on 24 July 2026 at the same rate — runs $5 per million input and $25 per million output, exactly half. Claude Sonnet 5 is $3 and $15. One pricing wrinkle worth knowing on Opus 5 specifically: its "Fast mode" runs at roughly 2.5x default speed for double the price, $10 and $50 — which happens to land at exactly the same rate as Fable 5, so don't confuse the two when comparing bills.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Fable 5 / Mythos 5 | $10.00 | $50.00 |
| Opus 5 (default speed) | $5.00 | $25.00 |
| Opus 5 (Fast mode) | $10.00 | $50.00 |
| Sonnet 5 | $3.00 | $15.00 |
| Haiku 4.5 | $1.00 | $5.00 |
There is no monthly subscription, no free tier, and no consumer app for these models. You pay per token through the API. If you're a student or hobbyist reading this wondering whether to switch, the honest answer is that Sonnet 5 or Haiku 4.5 will serve you better for a fraction of the cost — and Anthropic's own guidance is that Fable 5 is for the hardest reasoning and long-horizon agentic work, not everyday tasks.
The safeguards, and what they mean in practice
This is the part most coverage skips, and it's the part most likely to affect you if you actually use Fable 5.
Fable 5 ships with classifiers that watch for requests touching cybersecurity, biology and chemistry, or attempts to extract the model's own reasoning. When one fires, the request doesn't fail — it falls back to Claude Opus 4.8, which answers instead. Anthropic says more than 95% of Fable sessions involve no fallback at all.
Read that number carefully in both directions. 95% is high, and for most workloads you'll never notice. But it also means roughly one session in twenty gets silently answered by a different, less capable model — and if your work sits near security tooling or life sciences, your personal rate will be considerably higher than the average. Benign work adjacent to those fields can trigger false positives.
If you're building on the API, this matters architecturally. A refused request returns a successful HTTP 200 with stop_reason: "refusal" rather than an error, so code that reads the response content without checking the stop reason will break in ways that look like nothing went wrong. Anthropic's own documentation recommends configuring an explicit fallback model rather than letting a refused request simply stop.
One more constraint worth knowing before you build
Fable 5 requires 30-day data retention and is not available under zero data retention. If your organisation is configured for ZDR — common in regulated industries — every Fable 5 request returns a 400 error regardless of how well-formed it is. This catches people out because the error looks like a request problem rather than an account-configuration problem.
Where you can actually run it
Fable 5 is generally available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. Mythos 5 is not — it's distributed only through Glasswing partners, with a broader trusted-access programme for cybersecurity and biology researchers planned but not open. The one exception, added 21 August 2026: Claude Enterprise customers can now reach Mythos 5's cybersecurity capability indirectly through Claude Security's vulnerability-scanning beta, without joining Glasswing.
Technically, both offer a 1 million token context window and up to 128,000 output tokens per request. Two behavioural differences matter if you're migrating from an earlier Claude model: thinking is always on and can't be disabled, and the raw chain of thought is never returned — you can request a readable summary, but not the underlying reasoning. Requests that try to extract that reasoning are one of the categories the classifiers watch for.
Should you use it?
For the vast majority of readers of this site — students, early-career developers, people building side projects — the answer is no, and not because it isn't good. It's that you'd be paying frontier prices for capability you won't exercise. Anthropic positions Fable 5 for demanding reasoning and long-horizon agentic work: tasks that run autonomously for extended periods, complex multi-step engineering, research-grade analysis. If your workload is "summarise this document" or "help me debug this function," a cheaper model produces equivalent results.
Where it earns its price is the genuinely hard end: long autonomous runs that would otherwise need human correction partway through, first-shot implementations of well-specified systems, work where a single failed run costs more than the token bill. If you're not sure whether your task qualifies, it probably doesn't — and that's a cheap thing to test on Sonnet 5 first.
If you do work in cybersecurity or biology and are hitting fallbacks on Fable 5, the intended path is Project Glasswing rather than prompt-engineering around the classifiers. That's a partnership process, not a signup form.
Frequently asked questions
Is "Claude Fable 5 Mythos" a storytelling tool?
No. It isn't a storytelling tool and it isn't a single product. Claude Fable 5 and Claude Mythos 5 are two separate general-purpose language models released by Anthropic on 9 June 2026, built for reasoning, software engineering, knowledge work, vision, and scientific research. The storytelling framing that circulates online — including in an earlier version of this article — appears to come from misreading the model names.
What's the difference between Fable 5 and Mythos 5?
They share the same underlying architecture and the same pricing. Mythos 5 is the same model as Fable 5 with safeguards lifted in some areas, and it is distributed only through Project Glasswing to vetted cybersecurity and infrastructure partners. Fable 5 is the version available to the general public.
How much does Claude Fable 5 cost?
$10 per million input tokens and $50 per million output tokens — the same for both models. There is no subscription tier and no free tier; access is per-token through the API. For comparison, Claude Opus 4.8 is $5 and $25, and Claude Sonnet 5 is $3 and $15.
Can I sign up for Claude Mythos 5?
Not for direct model access — Project Glasswing remains a vetted-partner programme, not a signup form, and Anthropic has said it plans to extend that route to biomedical researchers but hasn't opened it generally. There is a narrower second path since 21 August 2026: Claude Enterprise customers can now enable Mythos 5-powered vulnerability scanning through Claude Security, a self-service beta activated from the admin console and billed as standard token usage. It's not the model itself — you get scan results with severity ratings and suggested patches, not a prompt box — but it's the first way to use Mythos 5's capability without going through Glasswing's vetting process.
Why did my Claude Fable 5 request get answered by a different model?
Fable 5 runs classifiers on incoming requests covering cybersecurity, biology and chemistry, and attempts to extract the model's internal reasoning. When one triggers, the request falls back to Claude Opus 4.8 rather than failing. Anthropic reports that more than 95% of sessions involve no fallback, but benign work adjacent to those fields can produce false positives.
Why do all my Fable 5 requests return a 400 error?
If every request fails regardless of how it's formed, check your organisation's data-retention setting before debugging the request itself. Fable 5 requires 30-day data retention and is not available to organisations configured for zero data retention — those accounts receive a 400 error on every request.
Is Claude Fable 5 worth it for a student or hobbyist?
Generally no. At $10/$50 per million tokens it is Anthropic's most expensive model, priced for demanding reasoning and long-horizon autonomous work. For learning, side projects, and everyday tasks, Claude Sonnet 5 ($3/$15) or Haiku 4.5 ($1/$5) deliver comparable results on that class of work at a fraction of the cost.