What we found We put MiniMax M3 head to head with Claude Opus 4.8 on 198 real user prompts and measured the four things that decide whether a model belongs in production: cost, speed, accuracy, and error rate. M3 is genuinely impressive. On the majority of everyday traffic it delivers frontier-class answers at a fraction of the cost and latency. MiniMax M3 vs Claude Opus 4.8 on 198 real prompts: about 8x cheaper per answer, about 1.6x faster (7.0s vs 11.3s median), and within one point of Opus on 70% of prompts. Accuracy holds on simple (77%) and medium (81%) prompts and dips on complex ones (40%), which is where a router escalates to a frontier model. The headline numbers: Cost: about 8 times cheaper per answer ($0.54 against $4.24 across the 198 prompts, at public list prices). Speed: about 1.6 times faster to a complete answer (7.0s against 11.3s median), even though M3 writes out its reasoning. Accuracy: within one point of Opus, on a five-point quality scale, on 70% of prompts, and a strong answer of 4 out of 5 or better on 56%. For a model at a tenth of the price to keep pace with the frontier on most real prompts, while answering faster, is a remarkable result. On the bulk of everyday assistant traffic, MiniMax M3 is simply the better economic choice. Where the gap opens, and why that is fine No efficient model matches a frontier model everywhere, and M3 is no exception. Its quality tracks how hard the prompt is: | Difficulty | Answers within 1 point of Opus | |---|---| | Simple | 77% | | Medium | 81% | | Complex | 40% | On simple and medium prompts, which are the bulk of real traffic, M3 stays within a point of Opus about four times out of five. On genuinely hard prompts, deep reasoning and long multi-step problems, the gap widens. That is not a knock on M3. It is the fundamental trade every efficient model makes, and it is the right trade for the majority of traffic. The only job left is to send the hard tail somewhere else. The catch: you cannot eyeball which prompts to trust Here is the part that makes a naive swap dangerous. If you replace Opus with M3 everywhere, roughly a quarter of your answers come back clearly worse, and the failures are not neatly confined to the prompts that look hard. Even simple prompts trip it up about one time in six, and you often cannot tell which in advance. A short factual question can miss while a longer one sails through. So you cannot safely capture M3's economics by blanket-swapping, and you cannot capture them by routing on a crude difficulty label either. You have to decide per prompt, after seeing the answer, whether M3 got it right. How Nadir captures M3's economics in production That per-prompt decision is exactly what Nadir is built to do. For every request, Nadir sends the prompt to the cheap, fast model first, checks the answer, and escalates to a frontier model only when the answer is not good enough. The design goal is simple: the large share of your traffic that M3 handles as well as the frontier runs on M3, capturing its cost and speed advantage, while the hard tail quietly escalates so your users never see the drop. The result you are aiming for is MiniMax M3's economics on most of your traffic and frontier quality where it matters, from one endpoint. Two lines of code, bring your own keys, and a savings dashboard that shows the real delta against always running the frontier model. We are making Nadir's routing M3-native Routing safely means grading the cheap answer well, and a grader is only as good as the models it has seen. So we have been extending Nadir's verifier to natively understand MiniMax M3's outputs, its reasoning-style answers and its formatting, so M3 routing is as sharp as it can be. MiniMax M3 is good enough to be a first-class cheap tier, and we are investing to make sure Nadir routes it that way. Ship MiniMax M3 today MiniMax M3 is fast, inexpensive, and frontier-class on most real work. Nadir is how you put it into production without betting your quality on the hard tail. Point your traffic at Nadir, keep your own keys, and let the router send each prompt to the model that will answer it best for the least cost. How we tested: 198 real prompts sampled from an open dataset of anonymized chat conversations, spanning factual questions, coding, math, creative writing, and everyday assistant tasks. Each prompt was answered by MiniMax M3 and Claude Opus 4.8, and the two answers were scored blind on quality. Cost uses public list prices; latency is the measured time to a complete answer. A directional comparison across one prompt set, not an exhaustive benchmark.