Anthropic releases Fable 5 with safety locks to unlock Mythos-level abilities

Author: Li Dan, Wall Street Insights

On the verge of an IPO, Anthropic unveils a major product.

On Tuesday, Eastern Time, Anthropic officially released Claude Fable 5 and simultaneously launched Claude Mythos 5. Both are based on the same underlying model architecture, with Fable 5 being the first Mythos-level model openly available to the public, while Mythos 5 retains more complete capabilities and is currently only accessible to a select few trusted institutions.

Anthropic states that Fable 5 has become the company's most powerful publicly released model to date, achieving industry-leading performance in software engineering, scientific research, financial analysis, visual understanding, and other fields. This means that the Mythos technology route, previously restricted due to its strong network security and cyberattack capabilities, is finally beginning to reach the mass market.

However, Anthropic has also strictly "downgraded" Fable 5: questions involving sensitive areas such as cybersecurity, biology, and chemistry will automatically switch to the less capable Claude Opus 4.8 to reduce misuse risks.

The pricing for Fable 5 and Mythos 5 is $10 per million input tokens and $50 per million output tokens. Anthropic points out that this price is less than half of the Claude Mythos preview version and states that the new model release marks a step forward toward the company's goal: to bring advanced AI capabilities to as many users as possible in a fast and safe manner.

For Anthropic, which is currently preparing for an IPO, this is not just a model upgrade but also a "show of strength" to the capital market, demonstrating its ability to develop cutting-edge AI while managing risks.

How powerful is Anthropic’s "most advanced general model" Fable 5?

According to Anthropic, Fable 5 is the first Mythos-level model that can be safely deployed publicly.

Anthropic claims that Fable 5 reaches industry-leading levels on nearly all benchmark tests, especially excelling in long-term, multi-step, high-complexity tasks. The company emphasizes: "The longer and more complex the task, the more obvious Fable 5’s advantages over other models."

Software development is one of its most prominent capabilities.

Payment giant Stripe discovered during early testing that Fable 5 completed a code migration in a large codebase containing 50 million lines of Ruby code in just one day, a task that would normally take a team two months to complete.

In finance and knowledge work, Anthropic cites third-party evaluations indicating that Fable 5 achieved the highest industry scores in advanced financial reasoning, document analysis, table and chart understanding, and other tasks. Trading firm IMC also stated that the model performed outstandingly in fact retrieval, root cause analysis, conceptual reasoning, and expected return analysis.

Its visual capabilities have also been significantly enhanced.

Anthropic states that Fable 5 can extract precise data from complex scientific charts and even reconstruct web application source code from screenshots. In an internal test, the model completed the game "Pokémon FireRed" solely through visual input, whereas previous versions of Claude required additional auxiliary tools to do so.

Additionally, Fable 5 has further improved in long context and continuous memory.

Anthropic says the model can maintain focus over tasks involving millions of tokens and utilize its own notes to continuously optimize output results.

Mythos 5 is more powerful, but the public still cannot use it directly

Alongside Fable 5, the highly anticipated Mythos 5 was also unveiled.

Anthropic revealed that both products are based on the same underlying model, but Mythos 5 has removed some safety restrictions, thus retaining more comprehensive cybersecurity and scientific research capabilities.

Anthropic plans to offer Mythos 5 only through a dedicated initiative called Project Glasswing, opening it to a small number of rigorously vetted institutions, including U.S. government agencies, critical infrastructure operators, and cybersecurity defense organizations.

Anthropic claims that the model possesses the "world’s strongest cybersecurity capabilities" and will gradually expand access through the "Trusted Access Program" in the future.

In the life sciences field, Anthropic also disclosed some notable experimental results.

The company states that Mythos 5 can independently complete protein design processes and has achieved or surpassed the level of professional researchers in some drug development tasks. The Anthropic research team also mentioned that a new hypothesis about a mechanism for E. coli proteins proposed by the model was later validated by independent laboratory research.

What has Mythos experienced to go from being too dangerous for public release to official commercial use?

The most attention-grabbing aspect of this release is that Anthropic has finally addressed the core issues that previously hindered Mythos from being publicly deployed.

Earlier this year, when Anthropic first showcased Mythos Preview, it caused a stir in the industry. The company unusually warned that the model could discover and exploit software vulnerabilities in major operating systems and browsers, demonstrating unprecedented capabilities in cyberattack and defense, and therefore decided not to open it to the public.

This decision is extremely rare in the AI industry.

Typically, AI companies prefer to rapidly expand model coverage, but Anthropic chose to restrict access proactively, allowing only a few partners to test. Previously, Project Glasswing covered about 50 organizations; recently, an additional 150 organizations were added, bringing the total close to 200.

To promote public release, Anthropic added multiple layers of safety mechanisms to Fable 5.

When the system detects sensitive requests related to cybersecurity, biology, chemistry, or model distillation, users are actually receiving responses from Claude Opus 4.8 instead of Fable 5 itself. The company states that these restrictions are triggered less than 5% of the time.

Anthropic also conducted over 1,000 hours of internal "jailbreak testing" to find ways to bypass safety restrictions. The company reports that testers were unable to find a universal cracking method capable of broadly bypassing the protections.

The key battle before the IPO: Anthropic aims to demonstrate that capability and safety can coexist

From a business perspective, the release of Fable 5 and Mythos 5 comes at a critical stage of Anthropic’s development.

At the end of May, Anthropic completed a $6.5 billion funding round, with a post-money valuation reaching $96.5 billion, surpassing OpenAI to become one of the highest-valued AI startups globally. Shortly after, in early June, the company secretly filed for an IPO, paving the way for a potential listing this fall.

A question the market has been watching is whether Anthropic can truly translate its leading research capabilities into scalable commercial products.

Although Mythos was previously considered one of the most advanced cybersecurity AIs globally, its deployment was limited due to safety risks, restricting its commercial value. Now, with the launch of Fable 5, Anthropic has found a compromise—retaining most of its advanced capabilities while meeting safety requirements through dynamic routing and permission controls.

This model holds significant importance for potential investors.

On one hand, it proves that Anthropic’s technological capabilities remain at the forefront of the industry; on the other, it showcases its unique advantages in model governance and risk management—qualities that regulators and capital markets are increasingly paying attention to.

Meanwhile, the company has also increased the pricing for its high-end models. Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, roughly double Opus 4.8’s prices, but Anthropic believes that higher efficiency can offset some costs.

As the IPO approaches, Anthropic is telling a new story to the capital market: not only does it possess some of the most powerful AI models, but it can also balance safety, regulation, and commercialization. And Fable 5 is the latest core product in this narrative.