🔮 Published June 9, 2026 — every claim in this article is sourced and verifiable. Key facts: Anthropic released Claude Fable 5 today, the first publicly available version of its Mythos model and the most capable model the company has ever shipped to the public. It scores about 80.3% on SWE-Bench Pro against roughly 69.2% for Claude Opus 4.8 and 58.6% for GPT-5.5, and is reported to be the first model to clear 90% on Hex's analytical benchmark. It is available immediately through the Claude API as model id claude-fable-5, on the Claude Platform, in Claude Code, in the mobile apps, and via AWS, Google Cloud, and Microsoft Foundry; through June 22 it is included on Pro, Max, Team, and Enterprise plans, after which usage credits may apply until capacity expands. API pricing is about $10 per million input tokens and $50 per million output tokens — roughly double Claude Opus 4.8, and far below the 5x-Opus pricing of the original Mythos preview. Fable ships with hard guardrails: in high-risk areas like cybersecurity, biology, chemistry, and model distillation it refuses and falls back to Claude Opus 4.8, with safety classifiers reportedly firing in under 5% of sessions. The underlying Mythos model, introduced in April 2026 through the restricted Project Glasswing program, autonomously discovered thousands of zero-day software vulnerabilities — including bugs that had survived in OpenBSD for 27 years — which is precisely why Anthropic kept it out of public hands until today.
For two months, the most powerful AI model Anthropic had built was something almost no one outside a guarded consortium could touch. It had a mythological name, a security-clearance feel, and a reputation that read more like a cyberweapon than a chatbot: a system that, left to work on its own, surfaced thousands of unknown software flaws across the operating systems and browsers most of the internet runs on. On June 9, 2026, that model — in a carefully fenced-in form — became something you can open on your phone. Anthropic is calling the public version Claude Fable 5, and it represents a genuinely new tier sitting above the Opus line. This article explains, in plain language, what Fable actually is, what changed between the locked-down Mythos and the version you can now use, where it is meaningfully better than Opus 4.8, the unusual safety design that makes the most capable Claude also the one that refuses the most, what it costs, and whether you should switch.
What Claude Fable 5 Actually Is
Claude Fable 5 is the public, safeguarded release of the model Anthropic developed under the codename Mythos. Fable and Mythos share the same underlying model — the difference is the brakes. Fable 5 is the everyday version with strong guardrails built in; Mythos 5 is the same intelligence with some of those protections removed, kept inside Project Glasswing for vetted cybersecurity professionals defending critical infrastructure. In other words, Anthropic did not water down the model for the public; it kept the capability and wrapped it in refusals. For most people the practical headline is simpler: Claude now has a new top model that is noticeably stronger at software engineering, dense knowledge work, and reading complex visual material than the Opus line that preceded it.
From Mythos to Fable: The Model Anthropic Was Afraid to Ship
The backstory is the most interesting part, and it explains the unusual launch. Anthropic introduced Claude Mythos Preview in April 2026 and immediately did something rare: instead of selling access, it largely withheld it. The model was strikingly good at computer-security work — capable of autonomously finding and chaining together previously unknown vulnerabilities, the kind of bugs that take human researchers months to uncover. Anthropic reported it surfaced thousands of zero-days, including flaws that had sat undetected in widely used open-source software for over a quarter century. That cuts both ways. The same skill that lets a defender patch a hole before attackers find it also describes, uncomfortably well, an offensive capability. So Anthropic launched Project Glasswing — a restricted program that put Mythos in the hands of critical-infrastructure partners to harden the world's most important software, while keeping it away from everyone else. Over the following weeks access widened to hundreds of organizations across roughly fifteen countries, all focused on defense. Fable 5 is the next step: the capability, opened to the public, with the dangerous edges fenced off.
The Benchmarks: Where Fable Sits Against Opus 4.8 and GPT-5.5
Benchmarks never tell the whole story, but the gap here is large enough to matter. On SWE-Bench Pro — a demanding test of fixing real issues in real codebases — Fable's reported score is well clear of both the previous Claude flagship and OpenAI's current top model, and Anthropic frames Fable as the first model to break a long-standing ceiling on Hex's analytical benchmark. Read these as direction and magnitude, not gospel.
| Model | SWE-Bench Pro (reported) | What it's for |
|---|---|---|
| Claude Fable 5 | ~80.3% | Hardest engineering, deep research, vision-heavy work |
| Claude Opus 4.8 | ~69.2% | Reliable daily flagship; the fallback for blocked queries |
| GPT-5.5 (OpenAI) | ~58.6% | OpenAI's current top general model |
What It's Actually Like to Use
The clearest signal of a capability jump is what a model does on messy, real work rather than tidy tests. The launch points to three kinds of tasks where Fable separates from Opus: large-scale code migration, reading hard visual material, and sustained work that has to remember its own progress.
- Codebase-scale engineering. Anthropic cites Fable compressing a migration of a roughly 50-million-line codebase into days rather than months. For most people the lesson is not the headline number — it is that Fable can hold an enormous, interconnected project in view and make consistent changes across it without losing the thread.
- Vision that reads, not just sees. Fable can pull precise figures out of a scientific chart and reconstruct a working web app from a single screenshot. If your work involves dense diagrams, dashboards, statements, or design mockups, this is the upgrade you will feel first.
- Memory that compounds. In an experiment that let the model keep notes in files as it played the game Slay the Spire, Fable reportedly improved about three times more than Opus 4.8 over the same run — a proxy for how well it learns from its own history on a long task instead of starting fresh each turn.
The Guardrails: The Most Capable Claude Also Refuses the Most
Here is the genuinely unusual design choice, and it is the heart of the story. Fable is Anthropic's strongest public model, yet it is deliberately built to say no in specific domains. When a request touches high-risk territory — offensive cybersecurity, biology, chemistry, or attempts to distill the model into a copy — Fable does not attempt a careful partial answer. It blocks the response and quietly hands the conversation to Claude Opus 4.8 instead, which answers within its normal, safer limits. Anthropic reports these classifiers fire in under 5% of sessions, so ordinary coding, writing, analysis, and research are unaffected. It is the opposite of the usual pattern, where the most powerful model is also the most permissive. With Fable, raw capability and the willingness to use it on dangerous tasks were split apart on purpose — and notably, the public launch arrived only days after Anthropic itself was publicly warning that frontier AI is getting dangerous. Whether you read that as responsible or as having it both ways, it is the defining feature of this release.
Pricing and How to Access It Today
Fable is positioned as a premium tier above Opus, but far cheaper than the locked-down Mythos preview, which was priced at roughly five times Opus. The subscription path is generous at launch and then tightens, so the next two weeks are the cheapest window to test it.
| How to access | Cost | Notes |
|---|---|---|
| Claude API (claude-fable-5) | ~$10 / 1M input, ~$50 / 1M output | Up to a 90% discount with prompt caching; ~2x Opus 4.8 rates |
| Pro / Max / Team / Enterprise | Included through June 22 | After June 23, usage credits may apply until capacity expands |
| Cloud marketplaces | Usage-based | Available on AWS, Google Cloud, and Microsoft Foundry |
Should You Switch? Fable vs Opus 4.8 for Real Work
For most people, most of the time, Opus 4.8 is still the sensible default — it is fast, reliable, and Anthropic specifically tuned it to be about four times less likely than its predecessor to let bugs in its own code slip through unflagged. Reach for Fable when the job is genuinely hard: a sprawling refactor across many files, a research synthesis that has to track dozens of moving parts, or visual material that other models mangle. Think of it the way a studio thinks about render time — you do not use the most expensive setting for every frame, only the ones that need it. And keep one habit regardless of model: for anything that touches money, law, security, or health, treat even Fable's output as a confident first draft to verify, not a final answer to trust.
The smartest way to use Fable this month is as a scalpel, not a default. Keep your everyday chats on Opus 4.8 or Sonnet, then switch to Fable only for the one task that genuinely stalled — the multi-file bug you cannot trace, the chart your other model misread, the migration that felt impossible. Because Fable costs roughly double per token, reserving it for the hard 10% gives you the capability jump without the bill, and the contrast makes it obvious which problems actually needed a frontier model and which never did.
You do not need to manage five separate AI subscriptions to work this way. LumiChats puts Claude (including the latest Opus and Sonnet), GPT-5.5, Gemini 3.5, DeepSeek V4, and 40+ more models behind a single ₹69/day pass (about $1/day) — so you can draft on one model, pressure-test the answer on another, and escalate only the genuinely hard problems to the most capable model available. For anyone who needs frontier power on the busy days and not every day, a single pay-as-you-go pass beats stacking monthly plans you will not fully use.
01Is Claude Fable 5 actually available to the public now?
Yes. As of June 9, 2026, Anthropic released Claude Fable 5 as the first publicly available version of its Mythos model. It is reachable through the Claude API as claude-fable-5, on the Claude Platform, in Claude Code, in the mobile apps, and via AWS, Google Cloud, and Microsoft Foundry. Through June 22 it is included on Pro, Max, Team, and Enterprise plans, after which usage credits may be required until capacity grows.
02What is the difference between Fable and Mythos?
They are the same base model with different safety settings. Fable 5 is the public version with strong guardrails. Mythos 5 is the same model with some protections removed and is reserved for vetted cybersecurity professionals through Project Glasswing. Anthropic kept the capability and changed the brakes, rather than shipping a weaker model to the public.
03Why does Fable refuse some requests?
Fable is built to block responses in high-risk areas — offensive cybersecurity, biology, chemistry, and attempts to copy the model — and to fall back to Claude Opus 4.8 instead. Anthropic reports these refusals trigger in under 5% of sessions, so normal coding, writing, analysis, and research are unaffected. It is a deliberate split between capability and the willingness to apply it to dangerous tasks.
04How much does Claude Fable 5 cost?
API pricing is reported at roughly $10 per million input tokens and $50 per million output tokens — about double Claude Opus 4.8, with up to a 90% discount when prompt caching applies. That is far cheaper than the original Mythos preview, which was priced at around five times Opus and was not available to the public at all.
05Should I use Fable or Opus 4.8 for everyday work?
Opus 4.8 remains the better default for most tasks: it is faster, cheaper, and was specifically tuned for reliability. Use Fable for the hardest jobs — large multi-file engineering, dense research synthesis, and complex visual material — where its extra capability earns its higher cost. For high-stakes output, verify either model's work rather than trusting it outright.
06Is Fable safe to use given the cyber-vulnerability headlines?
The headline capability — autonomously finding software vulnerabilities — lives in the locked-down Mythos branch used by defenders, not in the public Fable model, which refuses offensive-security tasks and routes them to Opus 4.8. For ordinary users, Fable behaves like a very strong, very capable assistant with a clear set of hard 'no' zones built in.
