MiniMax M3 vs GPT-6 Luna
Too close to call. On the work we tested, MiniMax M3 and GPT-6 Luna can't be told apart. GPT-6 Luna is the cheaper one, by 9.8×.
GrindBench-PubSec
CMiniMax M3MiniMax52% 47–58%0.72¢ a task · 300 tasksCGPT-6 LunaOpenAI54% 49–60%0.11¢ a task · 300 tasks
Too close to call here: their ranges overlap, so on another set of tasks the order could flip.
| Kind of work | MiniMax M3 | GPT-6 Luna |
|---|---|---|
| ATIP redaction | 2% | 2% |
| Procurement evaluation | 68% | 69% |
| Performance analysis | 49% | 54% |
| Bilingual equivalence | 92% | 90% |
GrindBench-Banking
DMiniMax M3MiniMax37% 25–48%4.1¢ a task · 60 tasksFGPT-6 LunaOpenAI18% 10–28%0.38¢ a task · 60 tasks
Too close to call here: their ranges overlap, so on another set of tasks the order could flip.
| Kind of work | MiniMax M3 | GPT-6 Luna |
|---|---|---|
| Lending & mortgages | 40% | 20% |
| AML & KYC | 40% | 20% |
| Deposits & payments | 40% | 20% |
| Complaints & conduct | 50% | 30% |
| Credit risk | 20% | 10% |
| Capital & liquidity | 30% | 10% |
Each kind of work here rests on 10 tasks per model, so a single row can easily be 20 points off. A lead of 10 points or more is in bold; read it as a hint, not a result.