MiniMax launches M3.1 Flash at $0.15 per million input tokens, weights later
MiniMax launched M3.1 Flash on October 19: $0.15 per million input tokens, $0.60 output. Artificial Analysis scores it 43, below DeepSeek V4.1 Flash 1017 (45).
MiniMax officially launched M3.1 Flash on October 19. The company describes it as a Flash-tier language model with a generational leap in agentic tasks, coding, computer use and cybersecurity. The API list price is $0.15 per million input tokens and $0.60 per million output tokens. Weights are promised “later”, with no date. The model is also the one that ran anonymously as “Space Bunny”, which MiniMax now confirms was an early version of M3.1 Flash.
Key points
- Price: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens, according to MiniMax’s October 19 blog. Caixin reports the same prices in yuan: 1, 4 and 0.1.
- Vendor benchmarks: 61.7 on OSWorld V2 and 29.6 on ALE, which MiniMax says rank first among Chinese models; 73.7% on DeepSWE v1.1; 82% on CyberGym.
- Independent check: Artificial Analysis (AA) scores the model 43 on its Intelligence Index, below DeepSeek V4.1 Flash 1017 at 45. Its AA-Omniscience hallucination rate is about 93% (The Decoder, citing AA), and an average index task costs $0.22.
Background
M3.1 Flash follows MiniMax M3. AA lists M3 at 29 on the same index, at $0.30 per million input tokens and $1.20 per million output tokens, with an average of $0.51 per task. The new list price is exactly half of M3’s. MiniMax also says overall usage cost for agentic tasks fell 70% against its predecessor, and that online inference is nearly three times faster. Caixin ties the 70% figure to M3. CEO Yan Junjie had stressed an “extreme cost-performance” route on the second-quarter earnings call, Caixin reports.
We first wrote about this model on September 28, when it topped the OpenRouter and OpenCode daily usage charts as Space Bunny (Space Bunny report). Our October 3 preview entry had neither a price nor a benchmark. Most coverage pits it against DeepSeek’s V4.1 Flash 1017, updated on October 17 (the model card entry).
The facts
The table sets vendor figures beside independent ones.
| Item | Figure | Source and basis |
|---|---|---|
| API list price, per million tokens | $0.15 input, $0.60 output, $0.015 cached input | MiniMax blog; Caixin gives 1, 4 and 0.1 yuan |
| Pricing structure | Flat rate, no input-length tiers | MiniMax blog (“flat pricing”); Caixin says M3 is tiered by input length |
| Context window | 1M tokens | AA model page |
| Output speed | 200+ tokens per second in everyday use; P50 170+ | MiniMax blog, vendor figure |
| OSWorld V2 / ALE | 61.7 / 29.6 | MiniMax blog; MiniMax says first among Chinese models |
| DeepSWE v1.1 / CyberGym | 73.7% / 82% | MiniMax blog |
| BankerToolBench | 76.2 | MiniMax blog |
| AA Intelligence Index | 43 | AA model page; M3 at 29; DeepSeek V4.1 Flash 1017 at 45 (AA leaderboard) |
| AA cost per task | $0.22 | AA model page; M3 at $0.51 |
| Tokens generated for the index | 180 million | AA model page; M3 at 120 million |
| AA-Omniscience | About 56% accuracy; about 93% hallucination rate | The Decoder, citing AA |
| Weights | “Later”, no date | MiniMax blog; The Decoder reports October 23, from a single source |
Vendor figures come from MiniMax’s October 19 blog. AA figures come from its model and leaderboard pages, read on October 21; the AA-Omniscience figures are The Decoder’s rounded numbers. The Decoder’s first-day report gave the AA index as 42; the page now shows 43.
The blog gives only a handful of scores. It says the model “ranks among the top performers” on AutomationBench-AA without giving a number, and the Pro version is also “later.”
What others say
Independent coverage falls into three groups: hands-on tests, leaderboard comparisons and a dispute over benchmark gaming.
In a hands-on test, Leiphone gave two setups the same hand-drawn campus plan and building-size table, and asked each to build an interactive 3D campus in a browser from scratch. MiniMax Code with M3.1 Flash took 55.4 minutes; OpenCode with DeepSeek V4.1 Flash took about 121 minutes. Model bills came to about 6.5 yuan each. Both met the main requirements. It is one test, not a measure of overall ability.
In leaderboard comparisons, The Decoder used AA data and concluded that neither model wins across the board. On first-day figures, DeepSeek scored 45 to MiniMax’s 42. MiniMax led on Humanity’s Last Exam (48% to 43%) and on AA-Omniscience accuracy (56% to 43%), while DeepSeek led on CritPt (24% to 22%). Leiphone’s DeepSWE figures favor DeepSeek on coding, at 77.0% to 73.7%.
On the gaming question, BlockBeats relays a test by the Chinese blogger 猫老板 using a private question bank. The model scored near full marks on old questions and near zero on new ones, with a median of 51.76, against 23.86 for M3 and 56.03 for DeepSeek V4.1 Flash 1017. That is one blogger’s test, and we found no public replication. Wccftech reports a sharper claim from an AI researcher: that MiniMax may have intercepted AA-Omniscience questions during M3’s evaluation, and that a modified question returned the original answer. The article describes its own account as educated conjecture and says it contacted AA. The version we read has no reply from AA.
Our take
Price is what changes first for developers, but it has to be weighed against volume. AA counted 180 million generated tokens for its index run, against 120 million for M3, and calls the model very verbose. A lower per-token price does not automatically mean a lower bill.
MiniMax’s 70% and the halved list price measure different things. The 70% covers overall usage cost on agentic tasks; the halved price is the list rate. Keep them apart.
The independent index is the other half. AA places the model below DeepSeek V4.1 Flash 1017, and its roughly 93% hallucination rate is high. The Decoder, citing AA, says the model tries to answer every test question rather than decline when uncertain. If your product relies on factual answers without retrieval or human review, test that before anything else. It matters more than the headline benchmarks.
Three things to watch: when the weights arrive (MiniMax says “later”; The Decoder gives October 23, from a single source), the Pro version (also “later”), and whether AA or MiniMax responds to the gaming questions.
How to try it
There are three ways in: the MiniMax API at platform.minimax.io, an M Plan subscription, and the MiniMax Code desktop app. The blog’s price table names the model MiniMax-M3.1-Flash; confirm the exact API call name in the platform docs. Until the weights are released, the model cannot be run locally.
Before committing, measure cost per task on your own workload, with the same prompts and a pass rate, rather than comparing list prices per million tokens.