MiniMax M3.1 Flash: $0.15 tokens. Price the whole agent loop.
MiniMax M3.1 Flash costs $0.15 in and $0.60 out per million tokens and claims 70% cheaper agent tasks. Measure cost per finished task, not per token.
MiniMax M3.1 Flash went live today at $0.15 per million input tokens and $0.60 per million output. The Shanghai lab says agent tasks now cost 70% less than on its previous model. Cheap tokens are nice. But an agent bill is tokens times steps times retries, and only one of those is on the price sheet.
What actually happened
Per MiniMax's announcement, dated October 19:
- Pricing: $0.15 input, $0.60 output, $0.015 cached input, per million tokens. One flat rate.
- Speed: about 3x faster online inference than before, with output above 200 tokens per second and a median above 170.
- Benchmarks (vendor-reported): 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 82% on CyberGym, 76.2 on BankerToolBench.
- Where: the MiniMax API, the M Plan subscription and the MiniMax Code desktop app.
- Weights: "later." The blog gives no date and no license.
The 70% figure is about agent-task cost, not list price. MiniMax does not publish the old rates on the page, so we cannot check the math. MiniMax also says a larger "Pro" model comes later.
Why MiniMax M3.1 Flash matters for your business
Every month a model gets cheaper, and every month somebody switches their agent over and finds the bill went up. A model that needs eight tool calls where the old one needed five is not cheaper at a lower token price. Neither is one that fails a task and retries.
What we do before any model swap in a client's agent:
- Measure cost per finished task. Run 50 real tasks from your logs. Divide total spend by tasks that passed. That is the only price that matters.
- Watch the step count. Log tool calls per task. A jump there wipes out a token discount fast.
- Wait on "open weights" until there is a license. "Weights later" with no license is not a self-hosting plan. Do not pick a model for a privacy promise that is not yet on paper.
- Check your buyer's rules. Some customers now ask where a model comes from. Know your answer before the swap, not after.
If M3.1 Flash passes your eval at a lower cost per task, take the savings. Just get that number from your own tasks, not from a benchmark table.
Key takeaways
- MiniMax M3.1 Flash: $0.15 input, $0.60 output, $0.015 cached, per million tokens
- MiniMax claims 70% lower agent-task cost and about 3x faster inference; benchmarks are self-reported
- Open weights are promised later, with no date or license yet
- Agent cost is tokens times steps times retries — measure cost per finished task
- Run your own 50-task eval before any model swap
Cheaper tokens, same bill? We build agents with cost-per-task tracking and swap-ready model configs, so a new price sheet is a test, not a rewrite. Run your agent costs through our calculator, or send us your logs.
Sources: MiniMax.
- #minimax
- #llm-pricing
- #ai-agents
- #model-evaluation
- #open-weights
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Mistral Timestral forecasts on a CPU. Your license is probably fine.
Mistral's Timestral time-series models top forecasting benchmarks, run on a CPU, and cap free use at $20M monthly revenue. Most small businesses clear it.
Read itMicrosoft calls AI agents insiders: start them at zero access
Microsoft's agentic insider push adds Dynamic Permissions and Dynamic Agent Risk. The lesson works without an E7 license: give AI agents zero standing access.
Read it