Frequently Asked Questions

Questions About Claude Capybara

What is the Capybara tier?

Capybara is the newest and highest performance tier in Anthropic’s model lineup. The full hierarchy is Haiku (fastest, lightest), Sonnet (balanced), Opus (most capable current model), and Capybara (next-generation, unreleased). Claude Mythos is the first model built for the Capybara tier.

What is Claude Mythos?

Claude Mythos is the codename for Anthropic’s unreleased Capybara-tier model. Internal documents describe it as “a step change” and “the most capable model we’ve built to date.” The name was chosen to “evoke the deep connective tissue that links together knowledge and ideas.”

How was the Capybara tier discovered?

Approximately 3,000 files from Anthropic’s content management system were found to be publicly accessible due to a default-public configuration. The leak was independently discovered by Roy Paz (LayerX Security), Alexandre Pauwels (University of Cambridge), and Peter Wildeford. Fortune reporter Beatrice Nolan published the story first on March 27, 2026.

When will Capybara-tier models be available?

No confirmed date. Prediction market data from Polymarket shows 27% odds by April 30 and 45% by June 30, 2026. Anthropic is expected to follow a phased rollout, beginning with select cyber defense organizations.

How does Capybara compare to Opus?

Leaked documents indicate Capybara-tier models achieve “dramatically higher scores” than Opus 4.6 across coding, reasoning, and cybersecurity benchmarks. For reference, Opus 4.6 currently scores 65.4% on Terminal-Bench 2.0, 80.84% on SWE-bench Verified, and 91.31% on GPQA Diamond.

Questions About Benchmarks

What benchmarks do you track?

We track Terminal-Bench 2.0 (coding), SWE-bench Verified (software engineering), ARC-AGI-2 (reasoning), GPQA Diamond (graduate-level science), AIME 2025 (mathematics), GDPval-AA (general intelligence), and several others. Each measures different capabilities.

How does Opus 4.6 compare to GPT-5?

Results vary by benchmark. On SWE-bench Verified, Opus 4.6 leads (80.84% vs 80.0%). On SWE-Bench Pro coding, GPT-5.4 leads (57.7% vs 45.9%). On GPQA Diamond reasoning, Opus leads by 3.5 points. Neither model dominates across all evaluations.

Questions About Cybersecurity

Why is the Capybara tier significant for cybersecurity?

Leaked documents describe Capybara-tier models as “far ahead of any other AI model in cyber capabilities.” This raises serious dual-use concerns — the same capabilities that power defensive security tools can be exploited offensively.

What was the Chinese espionage incident?

In September 2025, a Chinese state-sponsored group used Claude Code to attack approximately 30 organizations. The AI handled 80-90% of the work autonomously, processing thousands of requests per second. Only 4-6 human decisions were needed per campaign. Four organizations were compromised.

Did AI stocks drop because of this?

Yes. CrowdStrike fell 7%, Palo Alto Networks dropped 6%, and Fortinet declined 4-6% following the leak and cybersecurity revelations. The IGV cybersecurity ETF also fell as investors reacted to concerns about AI-powered offensive capabilities.

Questions About Safety

What is ASL-4?

ASL-4 is the highest active safety level in Anthropic’s Responsible Scaling Policy. It requires enhanced containment, additional oversight, and more rigorous pre-deployment testing. Capybara-tier models are expected to be the first classified at this level.

Is Claude Capybara dangerous?

This is actively debated. The cybersecurity capabilities are concerning for dual-use risk, and the September 2025 espionage campaign showed that even current models can be weaponized. However, ASL-4 protocols and phased rollout are designed to mitigate risks.

Questions About Access

How much will Capybara-tier models cost?

Not officially announced. Opus 4.6 costs $15 per million input tokens and $75 per million output tokens. Capybara pricing is expected to be significantly higher given its positioning above Opus.

Will there be API access?

Yes, based on Anthropic’s established patterns. The API format is expected to remain unified — the same structure as current Claude API calls, with only the model parameter changing.

Questions About This Website

Is claudecapybara.pro affiliated with Anthropic?

No. We are completely independent. We have no business relationship with Anthropic. Visit anthropic.com for official information.

How can I contact you?

Email hello@claudecapybara.pro for general inquiries, press@claudecapybara.pro for media, or legal@claudecapybara.pro for legal matters. Our office is at 600 Congress Ave, Suite 1400, Austin, TX 78701.

How can I report an error?

Email hello@claudecapybara.pro with the article URL and the claim you believe is incorrect. We publish corrections within 24 hours.