Anthropic spent two months deciding whether the world was ready for this. The same model that mapped vulnerabilities across every major operating system and web browser. The same model that’s now compressing months of engineering into days. They built safeguards, ran 1,000+ hours of external bug bounty testing, and partnered with Amazon, Apple, Google, Microsoft, Cisco, and JPMorgan Chase just to handle it safely. Then they released it anyway.
Claude Fable 5 launched June 9, 2026 as Anthropic’s first Mythos-class model made available to the general public. Here’s what you actually need to know before you use it or decide not to.
- Claude Fable 5 is the most capable AI model Anthropic has ever shipped publicly, sitting a full tier above the Opus line with state-of-the-art scores in coding, vision, knowledge work, and scientific research.
- Best for developers, researchers, and enterprises running long-horizon, complex tasks; skip it if your work is straightforward chat or simple document summaries Opus 4.8 will cost less and perform comparably.
- The key thing to know: queries in high-risk areas like cybersecurity, biology, chemistry, and distillation are automatically routed to Opus 4.8, and you won’t be charged Fable prices for those rerouted requests.
- Biggest mistake to avoid: assuming the free access window lasts. Through June 22, 2026, Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans on June 23, Anthropic pulls it and requires usage credits going forward.
- If you need the uncensored version with safeguards lifted in cybersecurity areas, that’s Claude Mythos 5, and it’s not available to you unless you’re already a Project Glasswing partner
What Claude Fable 5 Actually Is (And Why It’s Different).
Most new AI model launches are incremental. A few points on a benchmark, some UX polish, a new context window. Claude Fable 5 is not that.
Two months after Anthropic rolled out Mythos Preview to a limited number of users citing concerns about the model’s potential to cause damage in the wrong hands the company said it’s ready to release an equally powerful model to the public. That’s the real story here. Fable 5 doesn’t come from the same family as Opus or Sonnet. The same underlying model as Claude Mythos 5, Fable 5 ships with safety classifiers that detect high-risk requests across offensive cybersecurity, biology and chemistry, and model distillation, routing those to Opus 4.8 instead.
So what you’re getting is Mythos-class intelligence with a set of locked doors. The doors cover a small percentage of use cases. For most developers and knowledge workers, they’ll never even notice the fallback.
Fable 5’s new AI classifiers flag dangerous requests and automatically route them to the weaker Claude Opus 4.8 model — and over 95% of sessions aren’t affected at all. The classifiers watch three main domains: cybersecurity, biology/chemistry, and distillation. Everything else? Full Mythos-class capability, no restrictions.
The reason Anthropic needed all this is worth understanding. When Mythos Preview launched in April 2026, it didn’t just score well on security benchmarks it proved particularly adept at finding vulnerabilities across every major operating system and web browser, despite not being designed for cybersecurity. That wasn’t a feature. That was an accident. And it’s precisely why Anthropic launched Project Glasswing before they released anything publicly.
The Benchmarks: What “State of the Art” Actually Means Here
You’ve seen benchmark tables before. Most of them are designed to make every new model look impressive. So let’s cut to what actually matters in practice.
Fable 5 posts 80.3% on SWE-Bench Pro, while the next-best model sits 11 points behind. For context, SWE-Bench Pro tests real GitHub issues from production codebases the kind of problems where “almost right” still means broken code. An 11-point gap on that benchmark is not noise. It’s a different class of model.
The coding capability is where this gets concrete fast. During early testing, Stripe reported that Fable 5 compressed months of engineering into days — in a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a single day that would otherwise have taken a whole team over two months by hand.
That’s not a benchmark. That’s a real engineering team watching their two-month roadmap collapse into a Tuesday.
On Cognition’s FrontierCode evaluation which tests whether models can pass difficult coding tasks while meeting the standards of high-quality production codebases Fable 5 scores highest among frontier models, even at medium effort. The “medium effort” qualifier is important. Most benchmark wins come from maxing out compute. Fable 5 is winning on efficiency.
Vision is the other standout. Fable 5 can extract precise numbers from detailed scientific figures and perform complex vision-based tasks like rebuilding a web app’s source code from screenshots alone. Previous Claude models needed elaborate helper systems just to navigate games like Pokémon FireRed. Fable 5 finished the entire game using only raw visual input no maps, no navigation aids, no game-state data fed in externally.
For knowledge workers doing finance, legal, and research work: On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 has the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving. IMC noted that Fable 5 aced their trading-analysis evaluations nearly across the board, including factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.
The Safety Architecture: How the Locked Doors Actually Work
This is the part most coverage glosses over. The safeguards aren’t just a content filter slapped on top. They’re structural.
Claude Fable 5 includes safeguards for cybersecurity and biology. Queries in these domains are automatically routed to Opus 4.8 if flagged and you won’t be charged Fable prices for rerouted requests. That last part is easy to miss: Anthropic isn’t just blocking the dangerous requests, they’re also not charging you for them at Fable rates. You pay Opus 4.8 pricing when the fallback fires.
The classifier system runs in real time. In high-risk areas like cybersecurity, biology, chemistry, and distillation, the model blocks responses and falls back to Claude Opus 4.8. The model also tells the user when it does this. Transparency matters here. You won’t silently get a weaker response without knowing why.
Anthropic conducted extensive testing to determine whether users can trick Fable 5 into providing answers it shouldn’t, and ran an external “bug bounty” program with more than 1,000 hours of testing, during which no one found a universal jailbreak for the model. That’s a meaningful data point. Jailbreak resistance at this scale, with this much adversarial testing, is not a given for frontier models.
One important requirement that trips people up: unlike other Claude models in GitHub Copilot, Claude Fable 5 requires data retention to operate Anthropic’s safety classifiers. This is a 30-day data retention requirement. If your organization has zero-retention policies for compliance reasons healthcare, finance, legal you need to check whether Fable 5 fits your data governance setup before building anything on it.
The honest truth about the safety system: it’s tuned conservatively by design. To release the model both safely and quickly, Anthropic tuned these safeguards conservatively they’ll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions. Anthropic openly admits there will be false positives. If you’re doing legitimate security research, for example, you may hit the fallback on questions that aren’t actually dangerous. That’s a real friction point for security teams. The fix a Trusted Access Program — is planned but not live yet for most users.
Fable 5 vs. Mythos 5: The One You Can’t Have
Let’s be direct about this. Fable 5 and Mythos 5 are the same underlying model. The difference is the locks.
Claude Mythos 5 will continue to be offered through Project Glasswing in partnership with the US government, replacing the earlier Claude Mythos Preview. Anthropic calls it the world’s strongest cybersecurity model it scored 78% on the ExploitBench benchmark, up from 69% for Mythos Preview and 40% for Opus 4.8.
Anthropic is also planning a Trusted Access Program for biology select researchers will get access to Fable 5 without biology and chemistry safeguards, though cyber safeguards will stay in place.
So the path to the uncensored version is:
- For cybersecurity: Project Glasswing partnership with the US government. Not a self-service sign-up.
- For biology/chemistry: Trusted Access Program, coming soon, application-based.
- For everyone else: Fable 5 as-is, with the 5% fallback rate.
The science capabilities locked inside Mythos 5 are genuinely extraordinary. Using Mythos 5, Anthropic’s internal protein design experts accelerated aspects of the drug design process by around ten times. The model executed all of the tasks normally completed by a scientist: choosing binding sites, selecting and running protein design tools, and recovering from failures along the way. Nine of 14 protein targets yielded strong candidates for drug design currently under investigation.
And this: Mythos 5 conducted novel genomics research in over a week of largely autonomous work it assembled single-cell data for millions of cells spanning 138 animal species and designed and trained a custom machine learning model to identify cells performing the same role in even distantly related organisms. With only high-level human input, Mythos 5’s trained model outperformed a recent model published in the journal Science, despite being 100 times smaller.
That’s not a chatbot. That’s a collaborator. And it’s locked behind a government-coordinated access program for a reason.
Pricing: The $50 Number in Context
The headline number that’s been circulating is $50 per million output tokens. That’s the API price. Here’s what it actually means for different types of users.
Claude Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, with the existing 90% input token discount for prompt caching. For comparison, Claude Opus 4.8 runs $5 input and $25 output. So Fable 5 doubles the price.
That sounds expensive until you factor in the efficiency gains. One early customer found Fable 5 completed a frontier physics research task in 36 hours using one-third the reasoning tokens it took GPT-5.5 four days to match. At $10/M input vs $5/M input, if you’re using three times fewer tokens, Fable 5’s effective cost on that class of task is lower. This efficiency advantage applies specifically to complex multi-step reasoning, not to simple or short-context queries.
The catch most subscribers will miss: the free window closes fast. Anthropic states that Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from launch through June 22, 2026. From June 23, usage requires usage credits unless Anthropic extends the window.
The subscription cut-off on June 23 has no firm timeline for reversal. Anthropic’s stated intent is to restore Fable 5 as a standard plan feature “when sufficient capacity allows” that’s an open-ended commitment. Don’t build workflows that depend on flat-rate access until that changes.
So your decision framework before June 22:
- API developers: Fable 5 is live now at $10/$50. Test it against your actual use cases this week.
- Pro/Max/Team subscribers: You have until June 22 to evaluate it for free. Use that window aggressively.
- Enterprise (seat-based): Same free window. Start piloting now, especially for complex coding or document-heavy work.
- Enterprise (consumption-based): Already works, billed at API rates from day one.
One more pricing note worth flagging: on Claude.ai subscription plans, Fable 5 counts as 2x usage. If you’re on a capped Pro plan and burning through your usage limit quickly, Fable 5 will drain it twice as fast. Budget accordingly.
Where Fable 5 Actually Lives (Access Points)
You can reach Claude Fable 5 through multiple surfaces right now.
Fable 5 is available to anyone through Anthropic’s Claude API and consumption-based Enterprise plans. The API model ID is claude-fable-5. On AWS Bedrock, it’s anthropic.claude-fable-5 with regional variants.
Claude Fable 5 is also available in GitHub Copilot for Copilot Pro+, Max, Business, and Enterprise users though note it’s currently off by default and requires admin policy to enable.
You can also access Fable 5 through the Claude Platform on AWS, Google Cloud, and Microsoft Foundry.
For developers already using AI agents and agentic frameworks, Fable 5 is particularly relevant because of how it handles long-horizon autonomous tasks. This isn’t a model you’d use for single-turn queries it’s designed for the kind ofagentic AI workflows where a model needs to plan, execute, recover from errors, and iterate across multiple sessions.
The 30-day data retention requirement applies regardless of which platform you access it through. That’s not a platform-level policy it’s a model-level requirement tied to how the safety classifiers work.
Who Should Use Fable 5 Right Now (And Who Should Wait)
Straight answer: Fable 5 is worth using immediately if your work involves any of these.
Complex, multi-day coding tasks. If you’re doing large-scale refactors, codebase migrations, or building systems that require sustained context across thousands of files, this is the first model that can credibly handle it without constant hand-holding. The Stripe example isn’t marketing it’s a real signal about where the capability threshold sits.
Document-heavy knowledge work. Finance teams, legal analysts, and researchers working with dense PDFs, charts, and tables will notice the vision improvements immediately. Previous models guessed at chart values. Fable 5 extracts precise numbers from complex scientific figures. For anyone doing due diligence, contract review, or research synthesis, that difference compounds quickly.
Long-context projects. If you’ve been frustrated by models losing track of context halfway through a large project, Fable 5’s memory improvements are material. The model stays focused across millions of tokens in long-running tasks and improves its outputs using its own notes. This is relevant for anyone using Claude for extended business automation workflows or customer service applications that need to maintain context across long sessions.
Who should wait or stick with Opus 4.8:
- Straightforward chat, Q&A, or single-turn tasks Opus 4.8 will do this just as well at half the price.
- Teams with strict zero-retention data policies the 30-day retention requirement is a hard blocker.
- Anyone on a capped Pro plan who doesn’t have usage credits the 2x usage burn is real.
- Security researchers doing legitimate offensive security work until the Trusted Access Program opens, you’ll hit the fallback constantly.
The comparison that matters for teams already evaluating alternatives: Fable 5 vs competing models like Grok or Gemini comes down to reliability and sustained reasoning quality on complex tasks. The 80.3% on SWE-Bench Pro is the clearest signal — that lead is large enough that it’s not a close call for serious engineering work.
The Bigger Picture: Why Anthropic’s Timing Is Deliberate
This launch didn’t happen in a vacuum.
Fable’s launch comes as Anthropic prepares to enter the public markets alongside OpenAI and Elon Musk’s SpaceX. It also follows the AI firm’s plea urging major global AI labs to establish a coordinated brake pedal on frontier AI development. Anthropic warned that systems are advancing so rapidly that they may soon achieve recursive self-improvement autonomously improving themselves without human intervention.
So Anthropic is simultaneously releasing the most powerful AI it’s ever made publicly available, warning the world that AI is advancing too fast, and preparing for a potentially massive IPO. All in the same week. That tension is real, and it’s not hypocrisy it reflects a genuine strategic bet: if Mythos-class AI is coming regardless, better to deploy it with safeguards than to cede the market to those who won’t bother.
The Project Glasswing model where Amazon, Apple, Google, Microsoft, Cisco, JPMorgan Chase and others got early access specifically to harden their own infrastructure against Mythos-level capabilities is actually a smart approach. Use the dangerous tool to defend against the dangerous tool, in a controlled environment, before general release.
For users who are evaluating AI tools for students or lighter use cases, Claude Fable 5 is probably overkill. But for anyone building production systems, doing research, or running serious engineering work, the capability jump is real and the pricing is workable if you’re efficient.
You have 12 days of free access on paid plans. Don’t waste them.
If you’re on Pro, Max, Team, or Enterprise: run your hardest, most complex task through Fable 5 this week. Not a simple test your actual worst-case workflow. A large codebase question, a 100-page document analysis, a multi-step research task. That’s the only way to know whether the capability jump justifies the usage credits after June 23.
API developers: test token efficiency on your specific use case before assuming Fable 5 costs more. On complex reasoning tasks, the token reduction can flip the math entirely. On simple tasks, it won’t stick with Opus 4.8 for those.
And if you work in biology or security research: watch for the Trusted Access Program announcement. That’s your path to the version that actually matches what Project Glasswing partners are using.
