🏷️
Internal Model Codenames
Tengu, Capybara, Fennec, Numbat — and what comes next
P2Meta / Governance
Summary
Tengu = Claude Code project codename. Capybara = new model family (Claude 4.6 variant, confirmed by Undercover Mode forbidden strings). Fennec = Opus 4.6. Numbat = unreleased model still in testing.
Technical Details
Opus 4.7 and Sonnet 4.8 confirmed in development (found in Undercover Mode's forbidden strings list).
Capybara reportedly comes in 'fast' and regular thinking tiers with a significantly larger context window. Internal benchmarks: Capybara v8 has 29-30% false claims rate — a regression from 16.7% in v4. This quantifies something the industry usually hides.
Bug fix revealed: Capybara can prematurely stop generating when prompt shape resembles a turn boundary. Mitigated with prompt-shape surgery, not model fix.
All rollout gated by tengu_* prefixed kill-switches for staged deployment and revert.
Official / Public Basis
Codenames found in source code forbidden strings list (undercover.ts) and feature flags. Opus 4.7 and Sonnet 4.8 confirmed in development. Capybara v8 benchmark data found.
Governance Concerns
Internal benchmarks showing 29-30% false claims rate in Capybara v8 (regression from 16.7% in v4) quantify something the industry usually hides. Transparency about model limitations should be a governance requirement.
LightHope Ecosystem Mapping
LightHope — understanding model selection strategy, performance regression awareness, how production AI systems handle model-specific quirks