Anthropic's leaked 'Mythos' model described as step-change in AI capabilities, pricing hints at 10x cost increase
Mar 27, 2026
Key Points
- Anthropic's unreleased Mythos model shows dramatic performance gains in coding, reasoning, and cybersecurity over Claude Opus, positioning it as a significant capability leap.
- Mythos pricing signals a 10x cost increase over current Claude rates, with token limits dropping for subscribers and a potential $1,000 monthly plan arriving.
- The leaked details risk triggering accelerated model releases from OpenAI and DeepMind, though price-sensitive users may resist higher costs despite capability improvements.
Summary
Anthropic is testing a new model called Mythos with select customers as a step change in AI capabilities. The model scores dramatically higher than Opus in coding, academic reasoning, and cybersecurity and is part of a new Capybara series designed to be larger and more intelligent.
Mythos pricing signals a major cost increase. Current Claude pricing runs around $10 per 1 million input tokens and $25 per 1 million output tokens. Leaked pricing shows Mythos at $0.005 per second, $100 per million input tokens, and $2.50 per million output tokens, roughly 10x the current rate. Subscription pricing remains nominally unchanged, but token limits are dropping, meaning subscribers effectively pay more per use. A $1,000 plan may be coming.
The model is not yet ready for general release, but the leak carries strategic weight. Competitors at OpenAI and DeepMind may accelerate their own model releases in response. Capability announcements tend to cascade through the market this way.
The pricing increase arrives as compute costs climb across the industry. Larger, more capable models demand more infrastructure, and that cost flows downstream to API users and subscribers. Some users already switch providers based on rate changes, suggesting willingness to pay more may have limits despite the capability jump.
Anthropmic has not publicly confirmed Mythos details. The information comes from leaked materials and customer previews.