This is an opinion column. Every few weeks a lab announces a lower price per million tokens, and every few weeks buyers conclude that their AI bill is about to shrink. The latest example is MiniMax’s M3.1 Flash. Our view is that the per-token price is one input to an agent’s cost and rarely the one that decides the total. Treat a price cut as a reason to run a test, not as a forecast.

What was announced

According to SiliconANGLE, MiniMax launched M3.1 Flash on October 19, 2026, with input priced at 15 cents per million tokens, output at 60 cents per million and cached input at 1.5 cents. The company says agentic tasks cost 70% less to run than on its previous model and that inference is close to three times as fast. The same report says the model is available through the MiniMax API and its subscription plan, that the company has promised to release the Flash weights later, and that the benchmark results quoted, including 61.7 on OSWorld V2 and 73.7% on DeepSWE v1.1, are the company’s own and have not been independently verified. The New Stack ran a companion piece with a pointed headline: the token price halved, “your AI agent bill may not follow.”

We are not disputing MiniMax’s numbers. The 70% figure is a vendor claim about the vendor’s own tasks, and the price list is simply the price list. Our point is about what buyers infer from them.

Why price per token is not price per task

An agent does not make one request. It loops: read the goal, plan, call a tool, read the result, decide again. Four things multiply the token bill regardless of the rate card.

Number of steps. A weaker or less careful model may need more steps to reach the same result, and every step is another full request.

Retries and recovery. When a tool call fails or the model produces an invalid action, the agent tries again, often with the error message appended to the context.

Re-reading context. Each step usually resends the growing conversation. Cached input helps, which is why cache pricing matters, but a long task still pays for the history repeatedly if the cache is missed or evicted.

Reasoning tokens. MiniMax’s documentation, as reported by SiliconANGLE, keeps reasoning on for every request with five effort levels. Reasoning is billed as output, and output is the expensive side of the price list.

A hypothetical worked example

The numbers below are invented for illustration and do not describe any real product or measured workload.

Imagine Model A at $0.30 per million input tokens and $1.20 per million output tokens, and Model B, half the price, at $0.15 and $0.60. Suppose a task takes Model A 10 steps, with an average of 20,000 input tokens and 1,000 output tokens per step. Input is 200,000 tokens and output 10,000 tokens, so the cost is $0.060 plus $0.012, or $0.072.

Now suppose Model B needs 30 steps because it makes more mistakes and retries, with context averaging 25,000 tokens and 1,200 output tokens per step. Input is 750,000 tokens, costing $0.1125, and output 36,000 tokens, costing $0.0216. The total is about $0.134, nearly double Model A’s, despite the half-price rate card. Flip the assumptions, so that Model B needs only 8 steps, and it is far cheaper. The lesson is not that cheap models are bad; it is that the answer lives in the step count, which only a trial on your own workload reveals.

What buyers should measure instead

Track cost per completed task, not per token. For a sample of real tasks, record total tokens in and out, number of steps, number of retries, cache hit rate and the success rate. Failed tasks that a human must redo should count against the cost. Latency matters too: a faster model that finishes in fewer wall-clock minutes can reduce infrastructure costs elsewhere. Our guide to cutting your AI API bill covers caching and model routing, and how AI pricing works explains the units on the price list.

What to watch next

  • Independent evaluations of M3.1 Flash on cost per completed task, not just accuracy.
  • Whether MiniMax releases the promised Flash weights, which would let others measure efficiency themselves.
  • Competing price moves, including from DeepSeek, whose off-peak Flash rate SiliconANGLE says matches MiniMax’s input price.
  • More vendors publishing steps-per-task and cache statistics alongside rate cards.

Key takeaways

  • Per-token price cuts lower the bill only if step count, retries and context size stay the same.
  • A claim such as “70% cheaper agentic tasks” is a vendor figure for its own workloads.
  • Output and reasoning tokens usually cost more than input tokens.
  • Caching can help a lot, but only when requests actually hit the cache.
  • Measure cost per successful task on your own work before switching models.

Frequently asked questions

Is the MiniMax price cut meaningless?

No. A lower rate card is real and may well reduce your costs. We argue only that you should verify the effect on your own tasks rather than assume it.

What is the single best number to track?

Cost per successfully completed task, including the cost of failures that need redoing. For the broader cost picture, see the economics of AI training and inference.

Do cheaper models always need more steps?

Not always. Some small models are very efficient on narrow tasks. The point is that step count varies by model and task, so it has to be measured.