The Blanket Swap

What we measured

MiniMax M3 lists at roughly one tenth of Claude Opus 4.8's price per token, so the question every team asks is simple: if you send your traffic to the cheaper model, what do you actually give up? We ran both models on 198 real user prompts and measured the four things that matter in production, cost savings, accuracy, error rate, and speed. Here is what we found.

Cost savings

At list prices MiniMax M3 is about ten times cheaper than Opus 4.8 per token. On our 198 prompts it cost about $0.54 against roughly $4.24 for Opus, close to eight times less. The per-answer saving is nearer 8x than 10x because MiniMax is a reasoning model that writes out its thinking, so it spends more output tokens per answer than Opus does. Either way, the cost difference is large and real.

Accuracy

To measure accuracy we had both models answer the same prompt and scored each answer on a five-point quality scale. MiniMax landed within one point of Opus on 70% of prompts overall, and produced a genuinely good answer (4 out of 5 or better) on 56% of them. How well it holds up depends heavily on how hard the prompt is:

On everyday questions MiniMax keeps pace with Opus about four times out of five. On genuinely hard prompts it falls well behind.

Error rate

An average hides the risk that actually matters: how often the cheaper model returns an answer a user would notice is clearly worse. We count that as a severe drop, two or more points below Opus. Across all prompts the severe-drop rate was 27%, and it more than triples with difficulty:

The takeaway is the shape of the risk. Even easy prompts fail severely about one time in six, and you cannot always tell which. A short factual question can trip it up while a longer, more involved one sails through, so the failures are not something you can eyeball in advance.

Latency

MiniMax M3 is also faster, not slower, despite writing out its reasoning. On the same prompts it returned a complete answer in a median of 7.0 seconds against 11.3 seconds for Opus, roughly 1.6 times quicker.

What this means for you

MiniMax M3 is a strong, fast, inexpensive model. On the majority of everyday prompts it stays within a point of Opus, costs about eight times less, and answers noticeably quicker. The catch is the tail: roughly a quarter of a wholesale swap comes back clearly worse, concentrated on harder work and rising to well over half on complex prompts. Replacing Opus everywhere is therefore a real quality risk. The safe way to capture the savings is to send easy, low-risk traffic to the cheaper model and keep hard traffic on the frontier model, rather than switching wholesale.

How Nadir captures M3's economics in production

That safe path is not a wholesale swap or a crude difficulty label, it is a per-prompt decision made after seeing the answer, and that is exactly what Nadir does. For every request it sends the prompt to the cheap, fast model first, checks the answer with a calibrated verifier, and escalates to a frontier model only when the answer is not good enough. The large share of 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. Two lines of code, bring your own keys, and a savings dashboard that shows the real delta against always running the frontier model.

We have also been extending Nadir's verifier to natively grade MiniMax M3's reasoning-style outputs and formatting, so M3 routing is as sharp as it can be. M3 is good enough to be a first-class cheap tier, and Nadir routes it that way.

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. This is a directional comparison across one prompt set, not an exhaustive benchmark.