MiniMax M3.1 Flash claims 3x speed and 70% cost cut for agent workloads

MiniMax lists its M3.1 Flash model with a blunt pitch: 3x faster and 70% cheaper, framed for agents. The listing went up on Product Hunt, and that framing is the whole story. A 70% price cut on a chat model is a marketing line. A 70% price cut on the model that sits inside an agent loop is a different kind of claim, because agents do not make one call. They make dozens, then hundreds, chained through tool use, retries, and self-checks. The unit economics of an agent product are set almost entirely by the model that fills its loop.

That is why the numbers matter more than the launch. Speed and price are the two variables that decide whether an agent product is viable at all. A model that is 3x faster cuts the wall-clock time a user waits between “go” and a finished task. A model that is 70% cheaper cuts the bill that arrives at the end of the month. Both are worth taking seriously, and both deserve the same caveat: these are vendor figures, published by the company selling the model, and Tessera has not independently benchmarked them. {/* TODO: verify the 3x and 70% figures against MiniMax’s own model card or pricing page */}

The agent loop is where the money goes

Consider what an agent actually does. It plans, calls a tool, reads the result, decides the result was wrong, tries again, summarizes, and hands back an answer. Every one of those steps is a model call. A simple coding task can burn ten calls. A research task that browses a dozen pages can burn fifty. Multiply that across every user and the inference bill stops looking like a line item and starts looking like the product’s biggest cost.

So a 70% reduction is not a discount. It is a change in what you can afford to build. Features that were too expensive to ship at the old price, like letting the agent retry three times before giving up, or running a verification pass on every output, become affordable at the new one. The cheaper model does not just make the same product cheaper. It makes a more ambitious product possible.

This is the part of the AI economy that gets less attention than training runs. The headlines go to the frontier labs and their giant clusters. The margin, though, lives at inference, and inference is a commodity market with brutal price competition. MiniMax is not the first to cut prices and will not be the last. The pattern is familiar from the last two years of Chinese model releases: aggressive pricing, fast iteration, and a willingness to undercut on cost to win developer adoption.

What “Flash” signals about the model tier

The name matters. “Flash” is the industry’s shorthand for a small, fast, cheap model that trades some capability for latency and cost. Google uses the same word for its Gemini Flash line. The tier exists because most production traffic does not need the biggest model. It needs a model that is good enough, fast enough, and cheap enough to run at volume.

The bet MiniMax is making is that agent workloads are the volume. If agents become the default way software talks to models, then the model that wins is not the smartest one. It is the one that is cheap and fast enough to be called fifty times per task without the bill exploding. That is a different competition from the one the frontier labs are running, and it rewards different things: inference efficiency, serving infrastructure, and price discipline rather than raw benchmark scores.

There is a real question about whether a cheaper, faster model can hold up inside an agent loop. Agents are unforgiving. A model that is 5% worse at following instructions can fail a multi-step task far more than 5% of the time, because errors compound across steps. A cheap model that needs more retries can end up costing more than an expensive one that gets it right the first time. The 70% headline assumes the model is good enough that you do not spend the savings on retries. That assumption is the whole ballgame, and it is the one number the listing does not give you.

The cheaper model does not just make the same product cheaper. It makes a more ambitious product possible.

The pricing war has a floor

MiniMax is a Chinese lab, and its pricing sits in a global market where Chinese model providers have spent two years pushing costs down faster than their Western counterparts. That pressure is good for builders and bad for anyone selling inference at a premium. The direction of travel is clear: the cost per token for a competent model keeps falling, and the models that do not fall with it get priced out of agent workloads.

The floor is set by compute. Every token still costs GPU time, and someone has to pay for the chips and the data centers behind them. A 70% price cut is only sustainable if the underlying serving is efficient enough to protect a margin, or if the company is willing to lose money to buy market share. Both happen in this market. Builders should read a price cut as an opportunity and a warning: the price you build against today may not be the price tomorrow, in either direction.

The practical takeaway for anyone building agents is to stop hardcoding a single model. The cost and latency of the model tier are moving too fast to bet a product on one vendor’s price sheet. Route by task, keep a cheap model for the steps that do not need reasoning, and hold the expensive model for the steps that do. MiniMax’s Flash is one more option in that routing table, and a 70% cut is a strong reason to test it.

What to watch is not the launch but the follow-through. If independent evaluations show M3.1 Flash holding instruction-following quality inside long agent loops, the price cut is real and the agent economy gets cheaper to run. If the retries eat the savings, the headline stays a headline. The number that will settle it is not 3x or 70%. It is how many calls a task takes before it succeeds.