Anthropic Unveils Fable 5.2 and Mythos 5.2 With Major Upgrades Ahead of IPO
Anthropic flexes its muscles ahead of its IPO. The publisher of Claude unveiled on Monday, November 9, the latest version of its most powerful model, available in two variants: Claude Fable 5.2 and Claude Mythos 5.2. The first, aimed at the general public, boasts greatly improved capabilities in coding, reasoning, vision, and speed. The second (Mythos) offers enhanced cybersecurity features for select organizations, through verification programs in cybersecurity and life sciences. Even more interestingly, both models gain new abilities, especially in scientific research and robotics. Pricing has also dropped slightly.
The world's best general-purpose model?
| Tested capability | Claude Fable 5.2 | Claude Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon | Benchmark |
|---|---|---|---|---|---|
| Agentic coding | 49.1% | 45.5% | 42.8% | 41.1% | FrontierCode 2.0 |
| Bug resolution | 69.9% | 62.3% | 65.5% | 61.2% | FrontierSWE v2 |
| Business workflow automation | 51.4% | 42.5% | 41.4% | 53.9% | AutomationBench |
| Autonomous scientific research | 74.6% | 58.7% | 68.1% | 64.8% | Terminal-Bench-Science 0.1 |
| Cybersecurity (Mythos 5.2 score) | 76.3% | 67.6% | 61.0% | NC | CyScenarioBench |
| Web search | 82.1% | 72.3% | NC | NC | WANDR (relaxed score) |
| Computer use | 89.3% | 87.2% | 79.5% | 81.5% | OSWorld 2.1 (offline set, partial) |
| Robotics | 64.9% | 39.4% | 29.8% | NC | RoboDojo Sim |
| Multidisciplinary reasoning | 74.2% | 67.7% | 57.2% | NC | Humanity's Last Exam (with tools) |
| Vision | 73.4% | 64.4% | 71.0% | 72.5% | Chartography (no tools) |
Fable 5.2 and Mythos 5.2 are actually the same model. The main difference lies in the safeguards applied to their outputs. Mythos can handle advanced cybersecurity and biology tasks, while Fable 5.2 enforces strict safeguards and will often refuse to answer. As a result, Anthropic mainly highlights the benchmarks for Fable 5.2, which will be the most widely distributed model for the general public.
With this model (both the restricted and unrestricted versions), Anthropic goes beyond a simple frontier model for code. The company positions it as a universal model suited for any complex task. It excels in agentic coding, with scores well above Claude Opus 5.5, already a benchmark leader (49.1% versus 45.5%). It's close behind Gemini 4 Carbon in business task automation (51.4% versus 53.9%) and stands out in both offensive and defensive cybersecurity (that's Mythos, not Fable). In vision (chart analysis) and computer use, it also leads, with a slight edge. For web search, the gap is clear: 82.1% versus 72.3% for Opus 5.5. In short, this is a model designed to handle any task, no matter the domain.
The real breakthrough, beyond its strong scores in code and reasoning benchmarks, is in scientific research and robotics. The model makes a real leap in robotic simulation, scoring 64.9% compared to 39.4% for Opus 5.5 and just 29.8% for GPT-6 Astra. It also stands out in autonomous scientific research (74.6% versus 68.1% for GPT-6 Astra). Fable 5.2 managed to independently solve two still-open physics questions about black holes. Mythos 5.2, for its part, orchestrated two open source models to predict the structure of AAV capsids (the shell of viruses used in gene therapy), with 28 successes out of 34. A team of 1,000 Fable 5.2 agents also verified the proof of Fermat's theorem by computer in 30 hours, while Opus 5.5 only completed 26% in 18 hours. Finally, in robotics, Anthropic claims its model can generalize to some situations not seen during training. This ability is starting to make some researchers question the long-term relevance of world models.
One last point—and it's a key metric—the model is up to 30% faster than Fable 5.1. Artificial Analysis measures about 98 tokens per second at maximum reasoning speed, ahead of Gemini 4 Carbon (77 tokens per second). The model is also more efficient, letting it complete a given task more quickly.
Lower prices... on paper
| Model | Input ($/M tokens) | Output ($/M tokens) | Cache (read) ($/M tokens) | Fast input mode ($/M tokens) | Fast output mode |
|---|---|---|---|---|---|
| Claude Fable 5.2 | $8.00 | $40.00 | $0.20 | $16.00 | $80.00 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | $8.00 | $40.00 |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 | $20.00 | $100.00 |
| Gemini 4 Carbon | $4.00 | $20.00 | $0.10 | NC | NC |
Anthropic says prices are down about 20% for both input and output compared to Claude Fable 5.1 ($10 and $50 per million tokens). However, the model is consistently more expensive on tested tasks, according to Artificial Analysis. As a result, even though GPT-6 Astra is more expensive on paper, it's actually cheaper in practice: about $6 per Artificial Analysis benchmark task, versus around $10 for Opus 5.5 and nearly $27 for Fable 5.2. So Claude Fable 5.2 should only be used for complex tasks where Opus 5.5 and other frontier models fall short. For standard coding tasks, it will drain your quota faster or cost more via the API. Also, Fable 5.2 can use up to 50% of the weekly usage limits on standard Pro and Team subscriptions.
A model designed for the IPO?
With Fable 5.2 and Mythos 5.2, Anthropic is playing on two fronts. First, its historical clients now have a cutting-edge model for the most complex use cases: orchestrating dozens of agents, autonomous research, R&D, and cyber. Second, with the addition of autonomous scientific research, advanced biology, and especially robotics, Anthropic is now targeting new industries. The goal, reading between the lines, is clear: diversify its client base to boost recurring revenue from high-growth sectors. The payoff? Stronger revenue growth prospects and a broader customer base, which could make the future ANT stock even more attractive to potential investors.
This article was automatically translated from French. Read the original version: https://www.journaldunet.com/intelligence-artificielle/1557819-fable-5-2-mythos-5-2-le-coup-de-force-d-anthropic-juste-avant-son-ipo/