Anthropic 推出 Claude Fable 5:史上最強模型分成「公開」與「管制」兩版,AI 安全治理進入新階段

On the evening of June 9th, an engineer friend who works on AI applications in Taipei sent me a message with just one sentence: "It has released the model in two versions, have you seen it?" I opened Anthropic's announcement and was stunned for a few seconds - the key point is really not about the scores. This company has split the strongest capability of the same generation into two versions: one called Claude Fable 5, which you and I can use; and the other called Claude Mythos 5, which is only available to specific units that have passed the review.

In the few years I've been observing trends, I've seen the release of many "historically strongest models". However, the practice of "dividing the same capability based on the user's identity to determine whether you can get the full version" is still unprecedented in its brazenness. The significance of this event far exceeds its score on a certain list.

Event Background

Anthropic is the company behind the Claude series models, which has been gaining weight in the enterprise and developer circles over the past year or so - many people write programs, do search and answer questions, and build AI agents, all of which are based on Claude. Just at the end of May, it completed a massive fundraising round, with a valuation of nearly $1 trillion (we discussed this in our previous article Anthropic's massive fundraising). With the money and computing power in place, the next step is to release the next-generation model.

This time, the released model is different. In the past, models were released with different sizes - "large, medium, and small" - with the same capabilities, but differing in speed and price. The division of Fable 5 and Mythos 5 is in another dimension: it's not divided by performance, but by risk. Fable 5 is a version that has undergone safety processing and is suitable for general purposes; Mythos 5 has relaxed some safety restrictions and is only open to security partners (Anthropic's "Project Glasswing" program) and a few biological researchers who have passed the review.

In other words, Anthropic has determined that the capability of this generation of models is so strong that "full opening" is risky, so it has withdrawn the most cutting-edge part and only given it to people it trusts and who really need it.

This Time's Key Points

According to Anthropic's official announcement and AWS's simultaneous release, the key information about this release is as follows:

• Dual-version release: Claude Fable 5 (public, general use) and Claude Mythos 5 (controlled, limited to security and biological research units that have passed the review).
• Double pricing: The API price of Fable 5 is $10 per million input tokens and $50 per million output tokens - roughly twice that of the previous flagship Opus 4.8. The upgrade in capability means that your wallet will also need to be upgraded.
• Limited-time free trial: From June 9th to June 22nd, Pro, Max, Team, and some enterprise plans can use Fable 5 for free; after June 23rd, they will need to use their credits. Enterprise plans that are billed based on usage will start being charged from the day they go online.
• Cloud synchronization: Fable 5 was launched on AWS on the day of its release, making it very convenient for teams that are already in the AWS ecosystem.
• Official benchmark: Anthropic's announcement mentions Cognition's FrontierCode program evaluation, stating that Fable 5 performs best among frontier models. As for the "80.3% on SWE-Bench Pro" figure that has been circulating, it needs to be clarified - this is from third-party testing institutions (such as Vellum) and not from Anthropic's official announcement.

I want to remind you of this last point. When a model is released, various score screenshots flood the internet, but what the officials say and what third-party testers say are two different things. Taking third-party numbers as official guarantees is one of the easiest pitfalls to fall into when reading AI news.

Market Impact Analysis

For general users in Taiwan, the most direct feeling is twofold: it's smarter, but also more expensive. If you are a Claude Pro or Max subscriber, make sure to take advantage of the free trial period before June 22nd to try Fable 5 - especially for difficult tasks that you feel ChatGPT or Gemini can't handle well, compare the differences. However, don't develop the habit of using the most expensive model for everything; use cheaper models for daily tasks and save Fable 5 for tasks that really need it.

For enterprise applications, the doubled pricing is a reality that needs to be taken into account. If a company integrates Claude into its customer service, document processing, and internal tools, and the call volume is high, the price of $50 per million output tokens will be felt in the monthly bill. My suggestion is: don't switch to Fable 5 all at once; instead, do a tiered approach - use cheaper models for simple tasks and only use Fable 5 for tasks that really require its capabilities. This is consistent with our previous observation when discussing Microsoft's self-developed MAI model: in 2026, the core of enterprise AI strategy is "mixing and matching multiple models and allocating tasks accordingly", rather than being loyal to a single vendor.

For developers, the "review before giving full capability" mechanism of Mythos 5 is a signal worth paying attention to. If you are developing security, penetration testing, or red team tools, and you want to get the "full firepower" of the model in the future, you may need to pass the supplier's qualification review first. This is a reasonable threshold for legitimate security research, but it also means that getting model capabilities is transitioning from "pay and get" to "qualified to get". When using Cursor or other development tools to access Fable 5, you are enjoying the capabilities of the public version; the part that really relaxes the restrictions is not something that anyone can get.

Future Development Trends

I believe that the true highlight of this release is that "AI security tiered governance" is starting to move from theory to product. In the past, discussions about AI security mostly stopped at white papers and principle declarations; Fable 5 and Mythos 5 have turned it into actual product decisions - using capability tiering and identity review, directly writing it into the release process.

It can be expected that other frontier model manufacturers will follow suit. When models become strong to a certain extent, the political and legal risks of "full openness" are too high, and "tiered release based on risk and identity" will slowly become an industry norm. For users, this means that the "strongest model" you get may actually be a "safety-castrated version" of the strongest model; for regulators, this provides a manageable interface.

Another trend is the decoupling of price and capability through membership systems. The free trial period, followed by the use of credits, and enterprise usage billing - Anthropic has played this game very finely. This approach will make users more accustomed to "paying extra for top-notch capabilities", and also make the division of labor between "cheap models for daily tasks" and "flagship models for difficult tasks" clearer.

TheAI Academy Summary and Commentary

Honestly, how high Fable 5's scores are will be surpassed by the next model in half a year, and that's not the point of this news. The point is that Anthropic has demonstrated the emerging game rule of "the stronger the capability, the more tiered the release" with this one release.

When the "strongest model" starts to determine how much it can give you based on your identity, what we're discussing is no longer just technology, but who is qualified to touch this fire.

The specific suggestions for Taiwanese readers are very practical: first, don't waste the free trial period before June 22nd, use Fable 5 to test your most difficult tasks and feel the value of the doubled price; second, if you or your company heavily relies on Claude, now is the time to do "model tiering" - allocate tasks to different priced models based on difficulty, and don't let the bill get out of control; third, developers who do security or agent development should start paying attention to the trend of "review before giving full capability", and understand the qualification thresholds in advance. The stronger the capability, the better, but how to use it smartly and affordably is the real skill.

Data Sources

• Anthropic official announcement: Claude Fable 5 and Mythos 5
• AWS News Blog: Claude Fable 5 on AWS
• CNBC: Anthropic launches Claude Fable 5 (2026/06/09)

(This article is compiled based on publicly available information, and the model pricing, availability, and functionality are subject to Anthropic's latest official announcements; third-party testing numbers in the article are marked with their sources, and readers should verify them before referencing.)