I was comparing models for a project and pulled up a pricing tracker site to check MiniMax M2.7's rate. It listed the model at $0.00 per million input tokens and $0.00 per million output tokens. For about ten seconds I genuinely considered whether MiniMax was running some kind of loss-leader promotion I hadn't heard about. It wasn't. That was a scraping error — a stale or broken data pull on the tracker's end, not a real price — and it's a good reminder that automated pricing aggregators are themselves just another layer that can be wrong, on top of the usual reseller-markup confusion.
Once I went to the actual sources — MiniMax's own documentation and a cross-section of the platforms reselling access to the same models — the real picture turned out to be more interesting than "here's a number," mostly because of how consistent most of it is, with a few platforms that clearly aren't playing the same game as everyone else.
The official rate card, direct from MiniMax
For MiniMax-M3, the flagship model, standard-tier requests with up to 512K input tokens are billed at $0.30 per million input tokens and $1.20 per million output tokens. That number comes with a detail worth knowing: MiniMax's own pricing page lists a struck-through $0.60/$2.40 list price with a "permanent 50% off" applied to reach the $0.30/$1.20 rate actually charged. Structurally that's just the price — a permanent discount is a price — but it's worth knowing the discount framing exists if you're ever comparing screenshots of MiniMax pricing from different points in time and see the higher number somewhere.
Cache reads are priced well below standard input — $0.06 per million tokens on M2.7, roughly a fifth of the standard rate — while cache writes cost more than standard input, at $0.375 per million. If your workload resends a large, mostly-static prefix on every call, that read discount matters more to your actual bill than the headline input rate.
Two mechanisms exist for paying for faster responses, and they're not the same across models — worth checking which one applies before you flip a switch expecting the other behavior. M3 offers a service_tier: priority setting billed at 1.5x the standard rate for preferential request admission and more reliable latency under load. M2.7 instead ships as a separate -highspeed model variant, same weights, different routing, priced at 2x standard. Setting service_tier: priority on a model that actually uses the -highspeed variant mechanism won't get you what you're expecting.
One more caveat from the docs worth flagging before you assume the full 1M-token context window bills at the same standard rate: inputs beyond 512K tokens sit in a separate, currently gated tier — limited availability, contact sales — rather than being automatically available at the standard rate. If your use case genuinely needs that upper range of the context window, that's worth confirming directly with MiniMax rather than assuming standard pricing extends all the way up.
The reseller landscape: unusually consistent, with a few outliers
Here's the part that actually surprised me. Pulling current per-million-token rates for MiniMax M3 and M2.7 across a dozen-plus resale platforms, most of them land on almost exactly the same number: $0.30 input / $1.20 output for M3, matching MiniMax's own standard rate almost to the cent, across platforms like novita, Together AI, tokenlab, fireworks, and ZenMux. A few — OpenRouter and DeepInfra among them — price M2.7 and M2.5 somewhat below that pack, closer to $0.15-0.25 input.
That consistency is genuinely different from what you'll find comparing Kimi K2 pricing across the same kind of platform list, where the spread is wide enough that platform choice meaningfully changes your bill. For MiniMax, most resellers appear to be passing through the model at close to cost rather than adding a meaningful markup, which flips the useful question: instead of "which platform is cheapest," it becomes "which platforms are charging noticeably more, and is there a reason."
And there are real outliers worth naming, because paying 2-4x the standard rate without realizing it is an easy mistake if you're not checking: a few platforms in the set I reviewed price M3 as high as $0.60-1.11 input / $2.40-4.60 output — two to nearly four times the going rate everyone else converges on. That's not automatically a ripoff; it could reflect a different SLA, dedicated capacity, or bundled features. But it's the kind of gap worth confirming you're getting something for, rather than assuming all resellers are pricing the same model at the same rate just because most of them are.
A cost estimate that accounts for MiniMax's actual mechanics
Generic per-token calculators miss the two things that actually move a MiniMax bill: the cache discount and the priority/highspeed multiplier, which work differently depending on the model.
// minimax-cost-estimate.js — accounts for cache discount and the priority/highspeed
// multiplier, which use different mechanisms depending on the model
const RATES = {
"m3": { input: 0.30, output: 1.20, cacheRead: 0.06, priorityMultiplier: 1.5 },
"m2.7": { input: 0.30, output: 1.20, cacheRead: 0.06, highspeedMultiplier: 2.0 },
};
function estimateCost(model, { inputTokens, outputTokens, cachedTokens = 0, fastMode = false }) {
const r = RATES[model];
if (!r) throw new Error(`Unknown model: ${model}`);
const regularInput = Math.max(inputTokens - cachedTokens, 0);
let inputCost = (regularInput / 1_000_000) * r.input + (cachedTokens / 1_000_000) * r.cacheRead;
let outputCost = (outputTokens / 1_000_000) * r.output;
if (fastMode) {
// M3: same model, priority service tier, 1.5x
// M2.7: a distinct -highspeed model variant, 2x
const multiplier = r.priorityMultiplier ?? r.highspeedMultiplier;
inputCost *= multiplier;
outputCost *= multiplier;
}
return +(inputCost + outputCost).toFixed(4);
}
// Example: same workload, standard vs fast, on each model
const usage = { inputTokens: 2_000_000, cachedTokens: 1_500_000, outputTokens: 300_000 };
console.log("M3 standard: $" + estimateCost("m3", usage));
console.log("M3 priority: $" + estimateCost("m3", { ...usage, fastMode: true }));
console.log("M2.7 standard:$" + estimateCost("m2.7", usage));
console.log("M2.7 highspeed:$" + estimateCost("m2.7", { ...usage, fastMode: true }));
Running a heavily cached workload like this through the numbers makes the cache discount's actual impact obvious — it's a bigger lever on the final bill than deciding between standard and priority/highspeed service in most cases, simply because so much of a repeat-context workload's tokens fall under the cheaper cache rate rather than the standard input rate.
What I'd actually check before committing
Confirm you're pricing the model and tier you'll actually use — priority and highspeed aren't the same mechanism and aren't interchangeable by name. Check whether your workload reuses context heavily enough that the cache rate matters more than the headline input number. If you're evaluating a reseller platform, compare its number against the roughly $0.30/$1.20 standard most of the market converges on for M3, and if it's notably higher, find out why rather than assuming markup is universal. And don't trust a single pricing tracker's number without a source check — mine told me a paid API was free, which should have been the first sign to look elsewhere.
Where a routing decision fits
If you're already calling other model families through a gateway and would rather not open a separate MiniMax account just for this one model line, that's a routing question rather than a pricing one — RouteAI, for instance, lists MiniMax's models in its catalog alongside DeepSeek, Qwen, Kimi, and GLM, which is worth checking against the direct-and-reseller numbers above the same way you'd check any other candidate, rather than assumed to be cheaper or pricier without confirming.
TL;DR: MiniMax's official M3 rate is $0.30/$1.20 per million tokens (a permanently-discounted price from a $0.60/$2.40 list rate), with cache reads at roughly a fifth of standard input cost and two different fast-response mechanisms (M3's 1.5x priority tier vs M2.7's 2x highspeed variant) that aren't interchangeable. Most resellers price within a few cents of that official rate — a handful charge 2-4x more, which is worth noticing rather than assuming away.
Website: https://www.fastrouteai.com
Top comments (0)