Anthropic's Claude Mythos AI model revealed in data leak, described as 'step change' in capabilities
What was leaked
A data leak from an unsecured, publicly-searchable data store revealed that Anthropic is developing and testing a new AI model called Claude Mythos. The leak included approximately 3,000 unpublished assets linked to Anthropic's blog, including what appeared to be a draft blog post announcing the new model.
Model details from the source
According to the leaked documents:
- The model is called "Claude Mythos" and is also referred to as "Capybara," which Anthropic describes as "a new name for a new tier of model: larger and more intelligent than our Opus models."
- Capybara represents a new tier above Opus in Anthropic's model hierarchy (which currently includes Opus as the largest/most capable, Sonnet as faster/cheaper, and Haiku as smallest/fastest).
- Compared to Claude Opus 4.6, Capybara gets "dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others."
- The draft blog post describes Claude Mythos as "by far the most powerful AI model we've ever developed."
- Anthropic considers this model "a step change and the most capable we've built to date."
Current status and rollout
The model is currently being trialed by "early access customers" as part of a cautious rollout strategy. According to the source material:
- The model is expensive to run and not yet ready for general release
- Anthropic is "being deliberate about how we release it" due to the strength of its capabilities
- The company is working with "a small group of early access customers to test the model"
- The leak also revealed details of a planned, invite-only CEO summit in Europe as part of Anthropic's drive to sell AI models to large corporate customers
Security implications
The draft blog post indicated that the company believes Claude Mythos "poses unprecedented cybersecurity risks." The leak itself resulted from what Anthropic described as "human error" in the configuration of its content management system, which made draft content publicly accessible.
Context for developers using AI coding agents
For developers who rely on AI coding assistants, this leak suggests significant improvements in coding capabilities may be coming from Anthropic. The specific mention of "dramatically higher scores on tests of software coding" indicates potential advancements that could affect tools and workflows that integrate with Claude's API.
📖 Read the full source: HN AI Agents
👀 See Also
KV Cache Architecture Evolution: From GPT-2 to Mamba
Analysis of KV cache memory costs shows GPT-2 used 300 KiB/token, Llama 3 reduced it to 128 KiB/token with grouped-query attention, and DeepSeek V3 achieved 68.6 KiB/token with multi-head latent attention. Mamba/SSMs eliminate KV cache entirely with fixed-size hidden states.
UW Researchers Plan to Use Teacher-Worn Cameras for AI Training, Parents Opt-Out
University of Washington researchers planned to have preschool teachers wear first-person cameras to record children for AI model training, with an opt-out consent model.
Anthropic Blocks Claude Subscriptions via Third-Party Tools
Anthropic has implemented server-side blocks on Claude Pro/Max subscriptions used through third-party OAuth integrations, citing subsidized access being taken advantage of at scale. The policy change includes 'Extra Usage' billing that makes these integrations economically unviable.
Developer Switches from Cursor Composer 2 and Kimi 2.6 to Qwen3.6:35b-a3b for Enterprise Workloads
A developer reports using Qwen3.6:35b-a3b for daily work on a 500-700k LOC enterprise suite, citing better performance than Kimi 2.6 and DeepSeek 4 Pro/Flash, with costs ~$0.08/1M tokens on OpenRouter.