Summary of Key Points
Anthropic recently released two AI models, Fable 5.1 and Mythos 5.1. Fable 5.1 is targeted at general users and enterprises, while Mythos 5.1 is available to specific research institutions. The main improvements focus on enhancing the performance of AI models for long-duration, continuous tasks (known as "Agents") and reducing the costs associated with such tasks. This is likely Anthropic's last major release before going public. The company has secretly submitted its IPO documents and is expected to start public offerings after September, with a valuation estimated in the trillion-dollar range. However, it faces challenges such as disputes over user quotas and high costs associated with computing power.
New Models: AI That Can Work Longer and More Efficiently
Fable 5.1 is not a complete overhaul but a targeted optimization:
- Performance Focus on Long-Duration Tasks: The model's performance in tasks such as scientific experiments and complex code processing has doubled (from 24.7% to 52.6% on the Terminal-Bench-Science benchmark), and it can run continuously for up to 38 hours without human intervention (e.g., automatically correcting experimental errors or running multiple experiments in parallel). However, there has been less improvement in general reasoning and code processing tasks, indicating that Fable 5.1 is designed for scenarios where AI needs to work continuously.
- Cost Reduction: The basic price remains the same (10 USD per million tokens for input and 50 USD for output), but the cost of "cache reading" has been reduced by 75% (from 1 USD to 0.25 USD). Cache reading refers to the ability to reuse previously processed system prompts and code libraries without re-computing them, which can save significant amounts of money for longer tasks. According to the company, the cost of complex Agent tasks can be reduced by up to 45%, although the basic price is still 2.5 times that of GPT. Developers have noted that a single complex code task could cost between 20 and 50 USD.
Mythos 5.1: The "Unlimited" AI Model for Research Institutions
Mythos 5.1 has the same capabilities as Fable 5.1, with the main difference being fewer security restrictions:
- It is only available to "verified institutions in the fields of cybersecurity and life sciences," allowing for high-risk vulnerability detection and protein design. For example, the protein binder designed by Mythos 5.1 has a hit rate of nearly 50%, compared to an industry average of 10-15%, which is 10 times better than the best solutions in competitions. It can also be used to generate high-resolution maps of Venus from NASA's existing data, improving the detail from 10 kilometers to 2 kilometers.
- The reason for the distinction is that general models may be misused for malicious purposes, while research institutions require more flexible and powerful capabilities. Anthropic manages these risks through a "trusted access program."
The Final Showcase Before Going Public
This release is Anthropic's final effort to impress investors before the IPO:
- IPO Timeline: The company has secretly submitted its IPO documents and is expected to release the prospectus after Labor Day in September, with an investor day in mid-September, and possibly going public by the end of September.
- Valuation Impressive but Cost-Intensive: The private valuation in May was 96.5 billion USD, and now investors expect the public valuation to exceed 2 trillion USD (some estimates are as high as 3 trillion USD, based on 30 times the expected revenue). Revenue growth has been impressive, with second-quarter revenue being 14 times that of the same period last year, and the company reported its first profit. Future revenue projections are even more ambitious, aiming for 190-200 billion USD by 2028.
- High Cost Expenses: Anthropic has just signed two major contracts for computing power: one with Nscale worth 45 billion USD over 6 years and another with Lambda worth 50 billion USD, totaling 80 billion USD (approximately 537.7 billion RMB). This means the company is betting heavily on future computing resources.
User Complaints: Quotas Increased, but Maybe Not in Real Terms?
The release of the new models has sparked controversy among users, who are concerned about their quotas:
- In August, Anthropic announced a 25% increase in standard weekly quotas starting September 14th, but this is compared to the old quota system from May. Users are currently using temporary quotas that are 50% higher than the old ones, so the actual quota in September will decrease from 150 to 125, a reduction of 17%. Users are criticizing the company for using semantic tricks and demanding more transparent and generous quota policies.
The Critical Question After Going Public: Can Revenue Cover Cost Expenses?
Anthropic's main challenge after going public is not whether its models are powerful enough but whether they can generate sufficient revenue to cover the high costs associated with computing power and other expenses:
- The company is targeting high-value use cases such as research (protein design), enterprises (financial and medical security), and long-duration Agent tasks, where customers are willing to pay. For example, reducing cache costs is aimed at encouraging enterprises to use AI for longer-duration tasks.
- However, the cost of computing power is a significant burden. With contracts worth 80 billion USD already in place, the company will need to continue investing in GPUs. Wall Street will be watching to see whether the model's capabilities can generate enough revenue to cover these expenses.
In summary, this news release is essentially an opportunity for Anthropic to showcase its capabilities and business potential before going public. The key question is whether it can turn its technological strengths into substantial financial success. While the company's models are cutting-edge, it also faces practical challenges, such as managing user expectations and controlling high costs.