--- title: "MiniMax says an agent task got 70% cheaper. The 17.9% gross margin is the hard part" date: 2026-11-03 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-minimax-m31-flash-task-economics language: en --- # MiniMax says an agent task got 70% cheaper. The 17.9% gross margin is the hard part > MiniMax priced M3.1 Flash at 1 yuan per million input tokens and says agent tasks cost 70% less — a per-task claim that only works beside a thin gross margin. Two documents landed on the same desk this autumn. One is a price list with numbers that look like a mistake. The other is an income statement with a gross margin that looks like a warning. Read together, they explain more about China's model business than any benchmark table. The price list belongs to MiniMax. The income statement covers the six months to June. ## Key takeaways - MiniMax M3.1 Flash lists ¥1 per million input tokens and ¥4 per million output, with cached reads at ¥0.1 — roughly US$0.15 / US$0.60 / US$0.015. - The company's real claim is not a cheaper token but a cheaper task: it says agent workloads cost 70% less than on the previous model. - MiniMax reported a 17.9% gross margin for the first half of 2026, up from 12.1% — thin for a software business, normal for one selling compute. - Growth did not stop the losses widening: revenue tripled, gross margin rose to 17.9%, and the adjusted net loss still grew 111%. - Cost per completed task, not cost per million tokens, is the number procurement teams should be demanding. ## What the price list actually says M3.1 Flash went live on 19 October. MiniMax lists three prices per million tokens: ¥1 for input, ¥4 for output, ¥0.1 for a cache read. The platform's dollar list is US$0.15, US$0.60 and US$0.015. Then come the operating claims. Online inference is "nearly three times" faster than the previous model. Day-to-day output runs above 200 tokens per second, with a median around 170. And the headline: across agentic tasks, the combined cost of use fell about 70% against the model it replaces. That last figure is doing a lot of work, and it is not the same kind of number as the first three. A price per token is a rate card. A 70% cost reduction on agent (智能体) tasks is a claim about how many tokens a task consumes, how many retries it avoids, and how much scaffolding the model needs to finish. It is a claim about the whole loop. ## Why the margin is the story MiniMax's 2026 interim results put first-half revenue at US$117 million (≈RMB 786M), up 283% year on year. Gross margin rose from 12.1% to 17.9%. Adjusted net loss widened 111% to US$293 million (≈RMB 2.1B). A 17.9% gross margin is not a software margin. It is closer to what a company earns when it resells something it has to buy — electricity, silicon, bandwidth. For a large model (大模型) vendor, inference cost (推理成本) is what that something costs. The shift matters more than the number. In the first half of 2025, consumer apps were roughly 70% of MiniMax's revenue. By the first half of 2026, the open platform and enterprise services line had grown more than sevenfold and reached 63.4% of the total. The company swapped a high-margin consumer business for a low-margin infrastructure business, because that is where the tokens are. Zhipu shows the same gravity. Its gross margin fell from about 50% to 26.4% as API revenue went from 15.2% to 86.5% of the total, largely because its high-margin on-premise deployment business shrank. This is the trade Chinese labs have made. Volume is easy to buy with price. Margin is not. ## The 200-trillion-token audition The most interesting detail is not on the price list. Before the model had a name, it ran under one. MiniMax says an early version of M3.1 Flash ran anonymously as "Space Bunny" and processed more than 200 trillion tokens. Developers chose it without knowing whose model it was, on price and speed alone. The company revealed the identity at launch. The stunt is a test benchmarks cannot run. Leaderboards measure a model under a fixed harness on a fixed task set. An anonymous deployment measures whether people keep paying after the novelty wears off. MiniMax is claiming the second kind of evidence, which is harder to fake. ## The benchmarks, and what they leave out The company reports a set of scores, all self-run: 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82% on CyberGym. Those names matter. BankerToolBench and OSWorld V2 measure agentic work, not question answering. DeepSWE measures software tasks end to end. CyberGym measures vulnerability discovery. This model is tuned for loops, not chat. That is also why the scores are fragile: agent results move with the harness. Change the retry policy, the tool schema, or the step budget, and a 61.7 becomes something else. MiniMax published scores without the harness, so the numbers are a direction, not a specification. ## Honest limitations The financial figures come from MiniMax's interim report as summarised by Chinese business media; I have not read the filing line by line, and definitions of gross margin vary. The 70% cost claim is the company's, measured on its own task mix; a team with different prompts may see another number. I have not run M3.1 Flash, and I have no independent measurement of its throughput or its real cost per completed task. The benchmark scores are vendor-run and unaccompanied by harness details. None of this is investment advice, and none of it predicts whether the strategy works: a 17.9% gross margin can improve through volume and cheaper serving, or deteriorate through a price war the company cannot win. Finally, this is one company's arithmetic. DeepSeek, Zhipu and Moonshot each have different cost structures and different exposure to consumer revenue, so the pattern is suggestive, not universal. ## What readers can do now - Replace cost per million tokens in your budget model with cost per completed task. Count retries, tool calls and abandoned runs, then divide. The ranking of vendors often changes. - If you are evaluating M3.1 Flash, run your own harness on twenty real tasks before the pricing discussion. The gap between demo and production lives in the harness, not the rate card. - When a vendor quotes a percentage cost reduction, ask against what baseline and on whose task mix. A 70% claim measured on someone else's workload is a marketing number. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-minimax-m31-flash-task-economics Free to quote with attribution and a link to the original.