Claude Fable 5 is an amazing AI model

Anthropic has just released its most powerful AI model ever to the public, the same technology that was previously deemed too dangerous for regular users. Now anyone can use it, at least until June 22. Here is everything you need to know about the model that can handle Pokémon on its own, migrate 50 million lines of code in a day, and made its own creators nervous.

Quick Facts

🚀 Launched on June 9, 2026 by Anthropic
🤖 First Mythos-classified model to be made publicly available
💻 80.3% on SWE-bench Pro, over 20 percentage points ahead of GPT-5.5 and Gemini 3.1 Pro
🔒 Built-in security filters for cybersecurity, biology, chemistry, and model distillation
💰 10 dollars per million input tokens and 50 dollars per million output tokens via API
⚡ Can operate autonomously for several days without human supervision
🎮 Completed Pokémon FireRed with only screen dumps as input
📅 Included free in Pro, Max, Teams, and Enterprise until June 22, 2026
🛡️ Same underlying model found hundreds of zero-day vulnerabilities in Firefox
📈 Requires 30 days of data storage for security monitoring, even for enterprise customers

What exactly is Claude Fable 5

On June 9, 2026, Anthropic did something they had never done before: they launched a single frontier model as two separate products. Claude Fable 5 is the new publicly available flagship, equipped with safety classifiers that send certain queries to Claude Opus 4.8. Claude Mythos 5 is the same underlying model with some of these protective measures removed, and it is not publicly available.

In simple terms: Fable 5 is Mythos for regular mortals. Mythos 5 is the real monster, available only to select organizations working with critical infrastructure. Think Batman and his non-lethal weapons policy, but for AI.

The name Fable comes from the Latin fabula, “what is told”, related to the Greek mythos. Anthropic has essentially named its two models the same thing in two different dead or semi-dead languages. The marketing department clearly had a good day.

The background of Project Glasswing and a model that frightened its creators

To understand why this launch is a big deal, one needs to go back to April 2026. Anthropic announced Mythos on April 7 and stated that they had no plans to release the model to the public. Instead, Project Glasswing was launched, where a consortium of companies used Mythos to find and address vulnerabilities in software.

Why the secrecy? Early tests showed that the AI was able to independently detect vulnerabilities in all major operating systems and web browsers. To manage the risks, Anthropic initiated Project Glasswing, a massive security collaboration with tech giants like Amazon Web Services, Apple, Google, Microsoft, and JPMorgan Chase, to identify and fix software bugs before a broader launch.

In other words: the model was so skilled at hacking that the creators felt compelled to first let it fix the internet before releasing it. It’s about like teaching a burglar to be a locksmith before giving him his own company.

Last week, Anthropic expanded access to hundreds of organizations in 15 countries, still focusing on organizations managing critical infrastructure. And now, with Fable 5, the rest of us finally gain access to the technology.

Performance that leaves competitors in the dust

Fable 5 is industry-leading on almost all tested AI benchmarks with exceptional performance in software development, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the greater the lead compared to other models.

And now to the numbers that make competitors sweat. On the SWE-bench Pro, Claude Fable 5 scored 80.3 percent, ahead of Claude Opus 4.8 at 69.2 percent, GPT-5.5 at 58.6 percent, and Gemini 3.1 Pro at 54.2 percent. On FrontierCode Diamond, Fable 5 got 29.3 percent, compared to 13.4 percent for Claude Opus 4.8 and 5.7 percent for GPT-5.5.

Read that again. On the weakest code benchmark, Fable 5 performs five times better than GPT-5.5. This isn’t a victory, it’s a humiliation.

The gap between Fable 5 and Anthropic’s own Opus 4.8, greater than 11 percentage points, is larger than the gap between Opus 4.8 and Google’s Gemini 3.1 Pro. Anthropic has not only beaten its competitors, they have outdone themselves by a larger margin.

In knowledge work, Fable 5 scored 1932 points on GDPval-AA, compared to 1890 for Claude Opus 4.8, 1769 for GPT-5.5, and 1314 for Gemini 3.1 Pro. On Humanity’s Last Exam, the model scored 59.0 percent without tools and 64.5 percent with tools.

Independent analysts confirm the picture. Nathan Lambert at Interconnects notes that Claude Fable 5 is definitely the smartest model available to the public, a remarkable leap on basically every relevant benchmark, at only double the price of the current Opus models.

Stripe and the code migration that shocked the industry

Benchmarks are one thing, reality is another. Here is the concrete example that caused the most jaws to drop. Stripe provided Fable 5 with a codebase of 50 million lines of Ruby and asked the model to carry out a codebase-wide migration. It was completed in a day. According to reports, the same work would have taken a developer team over two months.

Two months of work for an entire team. One day for an AI. That’s the kind of statistic that either gets you cheering or updating your CV, depending on which side of the screen you are on.

Agent work and the real superpower

What truly distinguishes Fable 5 from earlier models is not that it responds faster to individual questions. It’s about endurance.

Fable 5 can be run in an agent-based environment such as Claude Code or Claude Managed Agents and work for days at a time, planning across different phases, delegating to sub-agents and controlling its own work.

Early tests from the code review company CodeRabbit show that Fable 5 is particularly strong for autonomous coding work when the instruction is incomplete and the agent needs to explore the environment before it can build. The model guides exploration well: first it learns the environment, then it identifies which files, tools, and constraints are available, and then it builds based on that picture.

In other words: give it a vague problem and go grab a cup of coffee. Or two. Or go on vacation.

The well-known AI researcher Ethan Mollick has also tested the model and notes that Fable consistently outperformed essentially every other public model he has used, by a significant margin, and that it produced some astonishing results.

Vision and the Pokémon venture

Fable 5 is the new flagship model for tasks involving vision. It can extract precise numbers from detailed scientific figures and perform complex vision-based tasks, such as reconstructing the source code of a web app purely from screenshots. Previous Claude models struggled with Pokémon FireRed even when given additional tools, but Fable 5 completed the entire game with a minimal, vision-driven interface.

An AI that can play a classic Pokémon game just by looking at the screen, just like a human. A human with improbably fast reflexes, perfect memory, and zero need for bathroom breaks, yet still.

The model also understands diagrams, tables, and graphs embedded in files and PDF documents, which improves document-heavy work in finance, law, analysis, and architecture. For anyone who has ever tried to get an AI to read a convoluted Excel table pasted as an image in a PDF, this is big news.

Science and research at a whole new level

Here it gets really interesting, and perhaps a bit eerie. Anthropic states that Mythos 5 generates new scientific hypotheses that are favored by researchers in about 80 percent of cases. The model conducted a week-long autonomous genomics study and outperformed a recently published specialized model despite being 100 times smaller. Mythos 5 is also said to have achieved ten times faster work in parts of the drug design process.

An AI that independently conducts genomics research for a week and arrives at better results than specialized systems. The Nobel Committee soon has to consider whether the prize can be awarded to something without a personal identification number.

Safety filters and the dark side

Here comes the interesting part. Why are there even safety filters?

It’s the same underlying architecture as Claude Mythos Preview, the model that during internal testing independently discovered and chained security holes in open-source code on a scale that worried its own creators.

If a user submits a high-risk question, such as how to make poison, the model blocks its response and reverts to Claude Opus 4.8, which provides a safe answer instead. The safety filters are expected to activate in less than five percent of sessions on average, and the user will not be charged Fable prices for redirected inquiries.

Good news then: if someone tries to use Fable 5 for something inappropriate, they instead receive a friendly response from Opus 4.8. It’s like hiring the world’s most dangerous security guard who also has instructions to politely decline and offer a cup of tea instead.

Moreover, Anthropic has thoroughly stress-tested the protections. The company ran an external bug bounty that produced no universal jailbreaks during over 1,000 hours of testing, and external red teaming organizations also failed to find any.

Data storage and the controversy no one talks about

However, there is a cost for all this security, and it’s not just about dollars. With the launch of Fable 5 and Mythos 5, Anthropic requires 30 days of storage for all traffic, including for corporate customers who previously had agreements on zero data storage. This requirement applies to everyone without exception and sets a new baseline for AI security practices at the API level.

For European companies with strict GDPR requirements, this is something to chew on before diving into the new model. Security monitoring in all its glory, but data storage that cannot be opted out of tends to make compliance departments wake up in the middle of the night, cold sweating.

Not everyone is wholly positive about the setup either. AI analyst Nathan Lambert writes that the unevenly implemented security policies that Anthropic has rolled out are becoming a classic cautionary tale about how narrow and self-fulfilling notions of security and control rarely work. A fable about Fable, if you will. The irony writes itself.

The timing and warning about self-improving AI

It gets really dark here, almost in a comedic way. The launch comes just days after Anthropic urged the world’s major AI labs to establish a coordinated brake on the development of frontier-AI. The company warned that the systems are developing so quickly that they may soon achieve recursive self-improvement, meaning they autonomously improve themselves without human involvement.

Anthropic’s own words on X last week: internal data shows that Claude is accelerating AI development, a potential path towards recursive self-improvement where AI autonomously builds a more capable successor. It is happening faster than they thought.

In summary: AI is developing dangerously fast, someone should really hit the brakes. Anyway, here is our most powerful model ever, please enjoy! It’s kind of like giving a speech about traffic safety and then handing out the keys to a Formula 1 car to everyone in the parking lot.

The timing is also not a coincidence from a business perspective. Claude Fable 5 comes just days after Anthropic announced that it had confidentially submitted its IPO prospectus to the U.S. Securities and Exchange Commission (SEC), paving the way for a potentially historic stock market listing. Anthropic stated in May that its revenue run rate had grown to $47 billion, up from about $10 billion in annual revenue last year.

Nothing sells stocks like an AI model that crushes all benchmarks. Except possibly an AI model that crushes all benchmarks and is also so dangerous that it needs a babysitter.

Price and availability with a deadline to keep an eye on

Until June 22, Fable 5 is included in Pro, Max, Teams, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic will remove Fable 5 from these plans, after which usage will require consumption credits, with plans to reinstate the model as a standard feature in the subscriptions as soon as capacity allows.

You have barely two weeks left to play with the toy for free. After that, it will be your wallet that decides.

Via the API, the model is called claude-fable-5. It supports a context window of 1 million tokens, up to 128,000 output tokens, and costs 10 dollars per million input tokens and 50 dollars per million output tokens. This is less than half the price of Claude Mythos Preview, but still the most expensive of the major AI models globally. In comparison, Gemini 3.1 Pro costs 2 dollars input and 12 dollars output per million tokens, about five times cheaper on the input side.

But as an analyst pointed out: if Fable 5 solves a task in an attempt where a cheaper model takes five, the math turns quite quickly. Fable 5 additionally beats Opus 4.8 on everyday spreadsheet tasks at all levels of effort, achieving it with fewer iterations, which translates to 25 to 30 percent faster runs.

“Claude Fable 5 feels materially different. In blind reviews, our lawyers found that its redlines matched or surpassed our current model every time.” Aveek Duttagupta, technical staff at Harvey

Fable 5 compared to competitors

Advantages and disadvantages of Claude Fable 5

Advantages:

• Industry-leading performance on almost all benchmarks, often by a wide margin
• Can operate autonomously for several days and tests its own work
• Understands and analyzes visual material, charts, and PDF files at a top level
• Lowest hallucination rate among frontier models according to independent comparisons
• Included temporarily at no additional cost in existing subscriptions
• Context window of 1 million tokens

Disadvantages:

• Most expensive of all major AI models, $50 per million output tokens
• Mandatory 30-day data retention, even for enterprise customers with zero retention agreements
• Security filters limit capabilities in cybersecurity, biology, and chemistry
• Free access in subscriptions ends June 23
• Token-intensive design that can lead to surprising bills

Common questions about Claude Fable 5

Is Claude Fable 5 the same as Mythos? Fable 5 and Mythos 5 share the same underlying model weights, but Fable 5 has security filters that limit capabilities in risky domains. Mythos 5 is only available to vetted organizations within Project Glasswing, including Apple, AWS, Microsoft, and CrowdStrike.

How long is Fable 5 free for subscribers? Until June 22, 2026, Fable 5 is included at no extra cost in Pro, Max, Teams, and seat-based Enterprise plans. From June 23, consumption credits will be required, until Anthropic reinstates the model as default when capacity allows.

How often are the security filters activated? According to Anthropic, protective measures are activated in less than five percent of sessions. At least 95 percent of sessions run entirely on the model’s own responses, and users are not charged Fable prices for redirected queries.

What does the data retention requirement mean? All traffic to Fable 5 is stored for 30 days for security monitoring, even for companies that previously had zero data retention agreements. This is a non-negotiable requirement and something GDPR-conscious organizations should consider.

Is Fable 5 worth the high price? For long, complex, and autonomous tasks, most indications are yes, as the model often solves in one attempt what cheaper models require multiple attempts for. For simple everyday queries, the difference compared to cheaper models is marginal.

Sources

https://www.anthropic.com/claude/fable