MiniMax M3.1 Flash: a cheaper agent and computer-use worker lane

Introduction

Cheap agent workers are no longer a single-vendor lane. DeepSeek refreshed `deepseek-flash` on October 17. Two days later MiniMax put a Flash-class model on a lower list price and leaned into computer use, SaaS automation, and coding. The useful question for builders is not whether MiniMax published a blog. It is whether M3.1 Flash completes the same tasks at a lower fully loaded cost, with tool traces that survive a security review.

Primary source: MiniMax M3.1 Flash announcement. Adjacent Bacancy reading: DeepSeek Flash-1017 Pareto update, Search and Fetch grounding, Reflection Beam, and late August routing.

What shipped

MiniMax also states that Pro, a higher-intelligence sibling, will arrive later. Do not route production traffic to an unpublished Pro ID. Pin the Flash API product that is live today and keep a rollback to the previous MiniMax worker or to `deepseek-flash`.

Vendor benchmarks that matter for agents

MiniMax publishes a capability story around four buckets. All figures below are vendor-reported. Independent evals should be run on the buyer’s own harness before any default-route change.

MiniMax says M3.1 Flash ranks first among Chinese models on OSWorld V2 and ALE. That comparison is MiniMax’s, not a third-party league table. For routing, treat those two numbers as a reason to include computer-use tasks in the bake-off, not as a reason to skip the bake-off.

What is technically significant

• Price is the headline. $0.15 / $0.60 undercuts many closed worker models and sits near DeepSeek Flash list rates while adding a cached-input line at $0.015.
• Speed is the second headline. 170+ TPS median changes how interactive coding agents feel, and it changes how many tool rounds fit inside a latency budget.
• The model is positioned as an actor, not a chatbot: AutomationBench, OSWorld, ALE, and BankerToolBench are all tool-and-environment suites.
• Weights are still closed. Self-host, VPC, and air-gapped buyers cannot treat this as a Hugging Face drop.

How it compares with the October 17 cheap lane

DeepSeek-V4.1-Flash-1017 remains the open-weight, same-price Pareto update on `deepseek-flash`, with MIT weights and Search/Fetch grounding APIs. M3.1 Flash is a closed API with a lower published output price and a computer-use pitch. They are not substitutes.

What it means for developers

• Add M3.1 Flash as a labeled worker in the router. Keep planner traffic on Opus 5.5, GPT-6 Astra, or the current closed default until the bake-off finishes.
• Log input, output, cached-input, and wall time per completed task. Token list price is not task cost.
• Include OSWorld-style and SaaS-connector tasks, not only SWE-style unit tests. MiniMax is selling action, not chat.
• Download MiniMax Code only if the desktop agent is in scope. API-only teams can ignore the desktop client.
• Watch the weights date. When MiniMax publishes them, re-open the self-host decision. Until then, assume egress to MiniMax.

What it means for businesses

• A 70% cost claim is a vendor estimate on MiniMax’s own task mix. Finance should require a two-week A/B on the company’s workflows.
• CyberGym 82% is a capability signal for security-engineering agents. It is not a license to run unscoped offensive testing. Keep those jobs in an approved lab with human review.
• Chinese-lab APIs remain a procurement and data-residency question. US federal civilian buyers now also have a DeepSeek FASCSA order to map; MiniMax is a different vendor, but the same inventory discipline applies.
• Cached-input pricing rewards stable system prompts and reused tool schemas. Unstable agent prompts will miss that discount.

Two-week routing pilot

1. 01
   Freeze three task packs

   Coding (DeepSWE-like), computer use (one OSWorld-style flow), and one internal SaaS workflow with a pass/fail rubric.
2. 02
   Call both cheap lanes

   M3.1 Flash and `deepseek-flash` (1017) with identical tools, timeouts, and max steps.
3. 03
   Price completed work

   Record tokens, cache hits, retries, and human-fix time. Rank by cost per accepted task, not tokens.
4. 04
   Gate computer use

   Keep OSWorld-class jobs in a VM or browser sandbox. Do not grant production credentials on the first week.
5. 05
   Decide a default

   Promote M3.1 Flash only where it wins on accepted-task cost and policy. Keep DeepSeek where weights or Search/Fetch matter.

Conclusion

M3.1 Flash is MiniMax’s bid for the high-speed, low-price agent worker. The API is live, the weights are not, and the interesting numbers sit on computer use and coding rather than chat. The correct response is a measured bake-off against Flash-1017, not a fleet-wide cutover.

Sources: minimax.io — minimax m31f ; api-docs.deepseek.com — news261017 ; platform.minimax.io