Head to head: OpenAI: GPT-5.6 Luna vs Grok 4.20 0309 Non Reasoning

This was a low-separation matchup, but OpenAI’s GPT-5.6 Luna did enough to edge Grok 4.20 0309 Non Reasoning where it counted. The scoreline flatters neither model: one clean task win, five ties, and a statistical verdict that says Luna is ahead—but only by a lean margin.

Most of this contest was defined by mutual non-performance. Of the six tasks, five were ties, and four of those ties came because both models missed the question. That matters: this was not Luna dominating the board, and it was not Grok hanging tough through strong parity. It was largely a stalemate built on shared failures. The only real separation came in **SimpleQA · Politics**, where **GPT-5.6 Luna got the exact-match answer right and Grok missed**. That single hit is the entire practical basis for Luna’s aggregate lead, and in a benchmark this small, that’s enough to matter. On **Science and technology**, both models split their performance the same way—one correct tie, one incorrect tie—so Grok never found a category where it could claw back ground. The aggregate score, **20.0 to 10.0**, makes Luna look comfortably ahead, but the task-level record tells the truer story: **1 win, 0 losses, 5 ties**. That’s why the statistical read is appropriately restrained at **75% confidence (lean)**. Luna was better in this sample, but not by the kind of margin that justifies chest-thumping. **Final call: GPT-5.6 Luna wins, narrowly but legitimately.** It was the only model here to convert a contested task into a point on the board, while Grok 4.20 0309 Non Reasoning never produced a category win of its own. That’s enough for the verdict—even if this head-to-head says as much about both models’ limits as it does about Luna’s edge.

Matchup powered by OpenRouter.