Anthropic's Claude Mythos 5: The First 10-Trillion-Parameter AI Model

Anthropic has quietly confirmed the existence of Claude Mythos 5, internally codenamed "Capybara," a model that represents a historic milestone in AI scaling: the first system publicly acknowledged to reach 10 trillion parameters. Described internally as "by far the most powerful AI model we have ever developed," Mythos 5 is designed to operate in high-stakes environments where previous frontier models have fallen short.

The model achieves its massive scale through a Mixture-of-Experts (MoE) architecture, where only roughly 800 billion to 1.2 trillion parameters are active during any given inference pass. This sparse activation approach makes the 10-trillion-parameter system computationally tractable, while preserving the full knowledge capacity of the complete parameter set. Training costs are estimated between $5 billion and $15 billion — a figure that underscores how far the compute frontier has shifted in just a few years.

Benchmark data remain sparse, but leaked community discussions suggest significant gains on coding and multi-step planning tests. One striking data point: Mythos 5 reportedly discovered a security vulnerability in a production system that had gone undetected for 27 years — missed by every human engineer and automated scanner that had previously reviewed it. The model supports context windows of 500K to 1 million tokens, enabling deep synthesis across enormous codebases or research corpora. Anthropic positions it as "far ahead of any other AI model in cyber capabilities," with particular strength in detecting and explaining exploitable flaws in complex systems.

Despite the fanfare, Mythos 5 is not publicly available. Anthropic is running a controlled early-access program called Project Glasswing, granting API access only to vetted partners focused exclusively on defensive cybersecurity work. The company has cited ongoing efficiency optimization as the reason for the limited rollout, though dual-use concerns around the model's offensive cyber potential are clearly part of the calculus. When it does reach general availability, pricing is expected to be substantially higher than Claude Opus 4 — reflecting both its capabilities and its compute demands.

The release reinforces that scaling laws continue to yield meaningful capability gains when paired with architectural innovations like MoE. It also raises pointed questions about what happens when the most powerful AI systems are locked behind controlled access programs rather than broad developer availability — and whether that tradeoff between capability and safety will define the next phase of frontier AI development.