MiniMax’s launch artwork for M3.1 Flash. Image: MiniMax
Chinese AI company MiniMax has released M3.1 Flash, a reasoning model that costs half as much to use as DeepSeek’s cheapest model, which was upgraded only three days earlier, while scoring almost as well in independent tests.
MiniMax charges $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015, according to its API price list. That is half the standard rate for its bigger M3 model, and half the $0.30 and $1.20 DeepSeek charges for V4.1 Flash 1017, which it released on October 17.
What MiniMax says it can do #
MiniMax says the model “delivers a generational leap in agentic tasks, coding, Computer Use and cybersecurity”, and that it runs “nearly three times faster” than before. It claims output speeds above 200 tokens a second in everyday use, with a median above 170, and says that “for agentic tasks, the model’s overall usage cost is 70% lower than its predecessor.” An early version had been running anonymously under the name “Space Bunny”, during which time MiniMax says it processed more than 200 trillion tokens.
The company’s own benchmark results include 73.7% on DeepSWE v1.1, a coding test, 61.7 on OSWorld V2, which measures how well a model can operate a computer, and 82% on the CyberGym security benchmark. On Agents’ Last Exam, a chart in the announcement puts M3.1 Flash at a 29.6% pass rate for an estimated $111 in total, against close to 40% for Anthropic’s Claude Opus 5.5 at more than ten times the cost. These are MiniMax’s figures and haven’t been independently checked.
What the independent tests show #
Artificial Analysis, which runs its own tests on AI models, gives M3.1 Flash a score of 43 on its Intelligence Index, up from 29 for MiniMax M3. DeepSeek V4.1 Flash 1017 scores 45, so DeepSeek still edges it on overall intelligence.
The gap is in the bill. Artificial Analysis puts the average cost of an index task at about $0.22 for M3.1 Flash, against $0.26 for DeepSeek’s model and $0.51 for M3. The saving is smaller than the price list suggests because MiniMax’s model is “very verbose” by Artificial Analysis’s measure, using about 180 million output tokens to get through the index.
The speed claims look less convincing so far. Artificial Analysis measured 122.8 output tokens a second, well short of the 170 MiniMax quotes as its median and slower than the 216.5 it recorded for DeepSeek. The model has a 1 million token context window, accepts text and images, and replies in text.
China’s price race keeps going #
The two releases, three days apart, show how hard China’s labs are now competing on price for everyday AI work such as coding agents and office automation, rather than on raw power alone. MiniMax pitches M3.1 Flash at business workflows across sales, marketing, finance and HR, and says a more capable Pro version will follow.
That pressure is landing on American labs too. Nvidia-backed Reflection gave away a 501 billion-parameter model last week explicitly to take on China’s best, while in Washington the government has ordered civilian agencies to remove DeepSeek, which shows the political limits on how far cheap Chinese models can spread.
How to try it #
M3.1 Flash is available now through the MiniMax API at platform.minimax.io, the M Plan subscription and MiniMax Code, the company’s desktop coding app. MiniMax says the model weights “will be released later” but hasn’t given a date or a licence in its announcement, and it doesn’t list any regional limits there. Developers who want to run it on their own hardware will have to wait.
Why it matters #
A model within a couple of points of DeepSeek’s best cheap option now costs half as much per token, a week after DeepSeek itself held its price. For developers running AI agents at scale, the cost of competent reasoning keeps falling, and China’s labs are the ones driving it down.
Sources: MiniMax announcement, MiniMax API pricing, Artificial Analysis: MiniMax M3.1 Flash, Artificial Analysis: DeepSeek V4.1 Flash 1017.