Anthropic Launches Fable 5 and Mythos 5, Restricts Access for High-Risk Users

In April, Anthropic released a preview of the Mythos model. The model was not made publicly available because, during testing, it demonstrated exceptional cybersecurity and vulnerability discovery capabilities, autonomously identifying thousands of critical zero-day vulnerabilities—posing a risk of malicious exploitation if leaked.

To confine this capability strictly to defensive scenarios, Anthropic launched Project Glasswing, granting access only to 12 core security partners—including Apple, Google, and Microsoft—and over 40 critical infrastructure providers, with all usage strictly controlled.

Locking it first before releasing—this very process generates buzz.

Two months later, at 00:00 Beijing Time on June 10, Anthropic officially launched Fable 5 and Mythos 5.

Based on benchmark scores, they achieved the highest ratings across nearly all tests, particularly standing out significantly from other models in software engineering and long-task execution.

But discussions around this model quickly went beyond just how powerful it was.

Fable 5 and Mythos 5 actually share the same underlying model. The difference lies in the fact that Fable 5 is designed for general users, while Mythos 5 remains restricted to trusted security partners. This is the first time in Anthropic’s product line that the same model is governed by two distinct sets of rules for two different audiences.

Moreover, ordinary users do not receive the full version. Anthropic has added a safety classifier outside Fable 5; if a request touches on sensitive areas such as cybersecurity, biochemistry, or model distillation, the system automatically switches to the less capable previous-generation model, Opus 4.8, for the response.

Pricing is also noteworthy: input $10, output $50 per million tokens, which the official team states is approximately twice that of Opus 4.8. Starting June 23, Fable 5 will be removed from subscription plans such as Pro and Max; users wishing to continue using it will need to consume additional credits.

Individually, this set of measures makes perfect sense: stronger capabilities require security restrictions, higher costs justify price increases, and sensitivity to risk calls for phased rollouts. But when viewed together, the signal they send becomes nuanced. This contrasts with the competitive logic of the past few years, where large model companies raced to accelerate development and maximize openness, each striving to reach as many users as possible. Anthropic, by contrast, has chosen to actively narrow its access points, turning restrictions into a core part of its product strategy.

So, is this so-called unprecedented "strongest model" being overhyped?

Regardless of classification, first let’s assess the actual strength.

Software engineering is the core highlight of Anthropic’s latest update. On the SWE-Bench Pro test, Fable 5 achieved a score of 80.3%. This test evaluates whether a model can locate bugs, understand context, and write functional fix code within real GitHub repositories. An 80.3% score means Fable 5 can solve four out of every five real-world engineering problems.

On the Terminal-Bench 2.1 leaderboard, Fable 5 scored 88.0%, surpassing OpenAI's Codex CLI. Notably, Fable 5 is a general-purpose model, while Codex CLI is a vertical tool designed specifically for programming scenarios—making the performance gap even more indicative of Fable 5’s programming capabilities.

But the true measure of the difference lies in FrontierCode Diamond, a test that evaluates whether the model’s generated code meets production-grade codebase quality standards. Fable 5 scored 29.3%, Opus 4.8 scored 13.4%, and GPT-5.5 scored just 5.7%. All three figures are from Anthropic’s official disclosures, with GPT-5.5’s result coming from its internal testing.

Over the past few years, AI models' ability to write code has continuously improved, but they have long been stuck at a bottleneck: the code runs, but it's hard to maintain; it passes benchmark tests, yet frequently encounters issues when deployed in real-world projects.

Fable 5's breakthrough on this dimension shows that Anthropic's latest upgrade isn't just about improving problem-solving ability—it's advancing the model toward actual engineering delivery.

Programmer Li Xia told "AIX Finance" that AI-generated code often suffers from contextual inconsistency—it can accurately understand requirements at first but tends to forget information during long-term tasks, leading to high maintenance costs later on.

In his view, Fable 5 shows significant improvement in logical consistency during long tasks, allowing similar coding tasks to be completed in a single pass with higher accuracy. However, compared to Opus 4.8, Fable 5 is slower in generation speed, has a longer reasoning chain, and exhibits reduced overall response speed.

Visual capabilities have also improved. Anthropic states that Fable 5 can extract precise numbers from complex scientific charts and directly reconstruct application source code from webpage screenshots. The company also demonstrated practical examples, showing that Fable 5 can complete the game Pokémon FireRed using only screenshots of the gameplay, without any auxiliary tools. Previous models required complex auxiliary systems to perform similar tasks.

Regarding long contexts and memory, the official statement is that the longer and more complex the task, the more pronounced Fable 5's advantages become.

In addition, life sciences is another key focus area. Anthropic revealed that its single-cell data analysis model, built on Mythos 5, covers 138 species and outperforms a recently published model in Science, while using only one percent of the parameters.

Looking solely at the benchmark score, the overall performance has indeed improved significantly.

Now let’s look beyond splitting.

Fable 5 is equipped with a safety classifier that automatically switches to Opus 4.8 to respond whenever a user request involves cybersecurity, biochemistry, or model distillation, and informs the user that a model downgrade has occurred. The official team states that over 95% of daily conversations do not trigger this switch, and tasks such as writing, programming, and analysis are largely unaffected. However, actual experience may vary depending on usage scenarios.

In actual use, this boundary is easily triggered. Li Xia mentioned that he wanted to experience Apple’s Siri AI features on a domestic Mac, which required modifying certain system-level serial number parameters, but Fable 5 directly blocked the operation. Currently, the classifier is set conservatively and may produce false positives; the official team says it will continue making adjustments.

However, Anthropic also disclosed another layer of restrictions: for requests related to large model development—such as building pre-training pipelines or designing distributed training infrastructure—the model will actively reduce output quality in the background without informing the user.

Overall, Fable 5 has made measurable improvements across key metrics, but the auto-degradation mechanism somewhat impacts the user experience.

Anthropic’s latest model upgrade packages the same underlying model into two products targeting two distinct user groups.

Mythos 5 remains within the framework of the Glass Wing Project and is accessible only to 12 core security partners, such as Apple, Google, and Microsoft, along with over 40 key infrastructure providers; restrictions in cybersecurity and biometrics have been lifted. Fable 5 is available to end-user subscribers.

Subsequently, Anthropic plans to open a trusted channel for vetted researchers, providing a version of Fable 5 without biological and chemical restrictions.

This involves an issue the large model industry has never faced before: when a model’s capabilities are sufficient to autonomously discover vulnerabilities, full-scale deployment becomes a risk.

Image source / Pexels

This explains why Fable 5 and Mythos 5 are split into two versions. In the past, model tiers were based on parameters, with clear capability gaps between sizes. Now, even models with the same parameters are tiered based on trust thresholds.

Independent developer Zhan Bo believes that, from a security perspective, this logic is sound. Opening up Mythos-level vulnerability discovery capabilities without restrictions to individuals would significantly lower the cost of attacks and make them prone to abuse for malicious purposes. Locking down access first and gradually granting trusted access is currently the most prudent approach.

But security only explains why layering is necessary. Another question is whether ordinary users who can use Fable 5 actually benefit from it.

Based on the benchmarks and case studies released by Anthropic, this round of upgrades prioritizes programmers and engineering teams.

Zhan Bo used Fable 5 to optimize his interpreted language project written in Rust. In the same web service scenario, he compared it with FastAPI (Python-based) and Hono (Node.js-based). The interpreter using Fable 5 consumed only 9.8 MB of memory, compared to 43.3 MB for FastAPI and 63.0 MB for Hono, while also leading in throughput and latency metrics.

He believes that the same task could be completed faster and with better results using Fable 5. Developers prioritize output quality over price; as long as the model significantly improves project outcomes, a higher cost is acceptable.

Li Xia also noted that for programmers, the quality of model-generated outputs is positively correlated with income. The higher the output quality, the more significant the efficiency gains and the greater the returns. Improvements in model capabilities can be directly translated into higher project quality and time savings, naturally leading to stronger willingness to pay.

But in a different use case, the conclusion would be different.

AI influencer Xu Zilong uses his daily routine as an example: his work consists of coding, data analysis, and writing papers. He believes that current large models have excessive coding capabilities but lack proficiency in text generation. For groups such as researchers, content creators, and professionals in law and finance, the core demand for AI centers on Chinese language comprehension, long-form writing, and document processing.

Zhan Bo noted that the current trend is not users selecting models, but models selecting users. Heavy programming users are being filtered out by high-end models, while casual, light users are being directed toward more cost-effective alternatives. This indicates that AI is no longer a universally accessible tool, but rather a system that progressively segments users based on their payment capacity and usage intensity.

Even the most powerful model is only worthwhile in the hands of those who need it.

Fable 5's API pricing is $10 per million input tokens and $50 per million output tokens, twice the cost of Opus 4.8 and the most expensive among globally available frontier models today.

Comparison of Major Large Model API Prices

But what truly matters is not that the price doubled, but that the payment method has changed.

After Fable 5 launches, Pro, Max, and Team subscription users can use it for free for two weeks. Starting June 23, the model will be removed from subscription plans, and continued use will require purchasing additional credits, billed at API rates. Usage of Fable 5 during the subscription period consumes twice the quota of Opus 4.8.

Offering a free trial before moving users off a subscription also sends a signal: shifting users from a fixed monthly fee to pay-as-you-go pricing. The advantage of a subscription is predictability—users know exactly how much they’ll spend each month. Pay-as-you-go benefits the platform more, as revenue scales with usage, removing any cap on user spending. In short, Anthropic is telling users: the most advanced technology shouldn’t be subscription-based in the first place.

The timing of this shift is also noteworthy. On June 1, Anthropic had just secretly filed its IPO prospectus with the SEC, valuing the company at $965 billion, with an aim to go public as early as October this year. From the beginning of the year through the end of May, Anthropic’s annualized revenue grew from $9 billion to $47 billion, with Claude Code contributing over $2.5 billion, and enterprise customers accounting for the vast majority of revenue.

A company preparing for its IPO needs to demonstrate to capital markets its revenue growth potential and pricing power. Separating the strongest model from a fixed subscription and guiding high-value users toward pay-as-you-go pricing is financially logical.

Meanwhile, domestic models are doing the exact opposite.

At the end of May, DeepSeek announced a permanent 75% price reduction for its V4-Pro API. Shortly after, Xiaomi followed suit with a staggering 99% price cut for MiMo-V2.5-Pro, bringing its pricing nearly in line with DeepSeek's.

On one side, DeepSeek is driving prices to the floor; on the other, Anthropic is raising the ceiling—different players are employing different strategies.

The price reduction of domestic models is partly due to the release of technological benefits, as changes in the underlying architecture have freed up resources that are being reinvested to capture a larger market.

Anthropic’s logic is precisely the opposite. It doesn’t need to compete for market share with low prices—its high pricing itself acts as a filter, retaining high-value users who are willing to pay for cutting-edge capabilities.

Xu Zilong believes that AI will become increasingly expensive in the future, because demand growth far outpaces supply growth—computing power involves electricity, chips, and model training, all of which expand too slowly. AI computing power will eventually become infrastructure, much like 5G; however, unlike 5G, the supply of computing power is far more constrained than bandwidth, leading to corresponding price increases.

From a business model perspective, the AI industry is developing a layered structure similar to the early days of cloud computing. At the bottom are highly standardized, universally accessible capabilities with prices approaching zero—monetized through volume. At the top are cutting-edge capabilities controlled by a few vendors, with pricing power concentrated and monetized through high per-customer value. While the prices of general capabilities will continue to decline, the premium for frontier capabilities will persist long-term.

From capability tiers to paid tiers, the AI industry is following the path taken by cloud computing. Cheaper models are becoming more abundant, while the best models are growing increasingly expensive.

At the respondent's request, Li Xia is a pseudonym in this article.

This article is from the WeChat public account "AIX Finance," authored by Lei Jing and edited by Jin Yufan.