GPT-6-luna Unlocks More Reasoning Tokens via API: ~18k Tokens Scores ~80.5% on Terminal-Bench
LysandreJik · x · 2026-10-07
Testing by evalstate shows GPT-6-luna delivers more reasoning tokens per effort level when accessed with an API key — worth knowing if you call it through different routes/providers. Using Codex CLI 0.160.0 with terminal-bench 2.1 (excluding 10x QEMU tasks): for tb-21 tasks, 10.5k output tokens is the point of diminishing returns; API max reaches 18,000 tokens at a score of 80.5%, just below Luna 5.6 in the same testing.
More from coding & agent
- David Khourshid ships @statelyai/graph, a fully type-safe open-source graph library with JSON graphs — DavidKPiano · 2026-10-07
- Skip the degree: a 10-repo GitHub roadmap to land a $200K AI engineer job in 90 days — mdancho84 · 2026-10-07
- Obsidian Starter Kit v5: one knowledge base, one AI, everything runs from there — dSebastien · 2026-10-07
- Building an AI Agent Is Easy — Designing One That Reasons, Remembers, and Runs Safely Isn't — goyalshaliniuk · 2026-10-07
- Jev: open-source PowerShell module turns AI model output into typed, structured decisions — dfinke · 2026-10-07
- Van Gogh Starry Night town generator adds real-time brushstroke flow fields — simonxxoo · 2026-10-07