A Chinese company says its new model has cut usage costs by 70%. But look at the price sheet: the per-million-token price has only been cut in half. The other 20 percentage points are hiding in the cache column.
On October 19, MiniMax launched M3.1 Flash. It has a 428-billion-parameter mixture-of-experts architecture, but activates only 23 billion parameters at a time. It supports a one-million-token context window and takes images and videos as input. These claims all come from the company; no one outside has reproduced them yet.
The price sheet, though, is clear. $0.15 per million input tokens. $0.60 for output. Cached input: $0.015. The previous generation, M3, was $0.30 and $1.20. So the unit price has been cut in half.
The official press release says costs are down 70%. Those missing 20 percentage points aren't in the prices. Not one of them.
I mulled it over for half the night. $0.15 and $0.015 are both in dollars, and there's a tenfold difference. That's a pretty big opening. Anyone who's run an agent knows where the money really goes in a task: the context that has to be stuffed in again every round—the codebase, tool definitions, logs from the first dozen or so steps. You feed it over and over, without changing a word, and pay full price each time.
The cache price singles that out and charges one-tenth. That's how the total bill comes down. Unit prices drop by 50%; the total bill drops by 70%. Both numbers are right, but they're talking about completely different things. Two separate things.
The vendor didn't lie when it said costs were down 70%. The trouble starts when readers take that to mean everything has become cheaper. What's cheaper is the unit price; what's being saved is the cost of repetition.
Ethereum did the same thing years ago. In the 2019 Istanbul upgrade, the cost of data carried in a transaction fell from 68 gas per byte to 16. The reason was straightforward: those bytes had to be stored permanently, but didn't need to be executed by every node. Then, in the 2024 Dencun upgrade, Ethereum gave rollup data its own lane: 128 KB per blob, with fees calculated separately from execution gas, rising and falling on their own. The same bytes went into two different buckets, and the prices could differ by dozens of times.
No one along the way shouted that things were cheaper. No one said costs had fallen by some percentage. They only changed how charges were calculated. The effect was the same. That's all there is to it.
So how exactly are those 70% savings achieved?
They come from recognizing the parts of an agent workflow that get read repeatedly and charging one-tenth for them. If you can identify which parts of your own workload are repetitive, you get that discount. If you can't, you pay full price—not a cent less. That's the logic.
That's why it can say in its press release that the cost per task is under five cents. Sell by the token, and customers will pull out a calculator. Sell by the task, and customers can't compare. The company ran its own evaluation of 152 tasks, with a bill of $111. That number checks out; no one's bothering to check the others.
A few things are still up in the air. The weights are supposed to be released on the 23rd, but there's no mention of a license. All the benchmark scores are self-reported; no third party has reproduced them. It scored 73.7 on DeepSWE, while DeepSeek's V4.1 Flash scored 74.2, half a point higher. On terminal tests, it scored 38.1, while Opus 5 scored 13.7 points higher. These are all listed in its own table. It's not hiding them.
One more thing: reasoning can't be turned off. Even simple questions have to be thought through first. There's another bill hidden in the money you save. Who pays it? No clear answer yet.
If you're running agents, don't focus on the unit price just yet. Take a look at how much unchanged material you feed into the context each round. That's the part you can cut.
That money has been on the bill all along.
$BTC $ARB $TAO
#AI #大模型 #以太坊 #比特币 #CostReduction
On October 19, MiniMax launched M3.1 Flash. It has a 428-billion-parameter mixture-of-experts architecture, but activates only 23 billion parameters at a time. It supports a one-million-token context window and takes images and videos as input. These claims all come from the company; no one outside has reproduced them yet.
The price sheet, though, is clear. $0.15 per million input tokens. $0.60 for output. Cached input: $0.015. The previous generation, M3, was $0.30 and $1.20. So the unit price has been cut in half.
The official press release says costs are down 70%. Those missing 20 percentage points aren't in the prices. Not one of them.
I mulled it over for half the night. $0.15 and $0.015 are both in dollars, and there's a tenfold difference. That's a pretty big opening. Anyone who's run an agent knows where the money really goes in a task: the context that has to be stuffed in again every round—the codebase, tool definitions, logs from the first dozen or so steps. You feed it over and over, without changing a word, and pay full price each time.
The cache price singles that out and charges one-tenth. That's how the total bill comes down. Unit prices drop by 50%; the total bill drops by 70%. Both numbers are right, but they're talking about completely different things. Two separate things.
The vendor didn't lie when it said costs were down 70%. The trouble starts when readers take that to mean everything has become cheaper. What's cheaper is the unit price; what's being saved is the cost of repetition.
Ethereum did the same thing years ago. In the 2019 Istanbul upgrade, the cost of data carried in a transaction fell from 68 gas per byte to 16. The reason was straightforward: those bytes had to be stored permanently, but didn't need to be executed by every node. Then, in the 2024 Dencun upgrade, Ethereum gave rollup data its own lane: 128 KB per blob, with fees calculated separately from execution gas, rising and falling on their own. The same bytes went into two different buckets, and the prices could differ by dozens of times.
No one along the way shouted that things were cheaper. No one said costs had fallen by some percentage. They only changed how charges were calculated. The effect was the same. That's all there is to it.
So how exactly are those 70% savings achieved?
They come from recognizing the parts of an agent workflow that get read repeatedly and charging one-tenth for them. If you can identify which parts of your own workload are repetitive, you get that discount. If you can't, you pay full price—not a cent less. That's the logic.
That's why it can say in its press release that the cost per task is under five cents. Sell by the token, and customers will pull out a calculator. Sell by the task, and customers can't compare. The company ran its own evaluation of 152 tasks, with a bill of $111. That number checks out; no one's bothering to check the others.
A few things are still up in the air. The weights are supposed to be released on the 23rd, but there's no mention of a license. All the benchmark scores are self-reported; no third party has reproduced them. It scored 73.7 on DeepSWE, while DeepSeek's V4.1 Flash scored 74.2, half a point higher. On terminal tests, it scored 38.1, while Opus 5 scored 13.7 points higher. These are all listed in its own table. It's not hiding them.
One more thing: reasoning can't be turned off. Even simple questions have to be thought through first. There's another bill hidden in the money you save. Who pays it? No clear answer yet.
If you're running agents, don't focus on the unit price just yet. Take a look at how much unchanged material you feed into the context each round. That's the part you can cut.
That money has been on the bill all along.
$BTC $ARB $TAO
#AI #大模型 #以太坊 #比特币 #CostReduction