Authors丨Wu Haiming, Li Na
Editor丨Li Na
Just as DeepSeek pushed the benchmark for domestic Flash cost-effectiveness a step forward, MiniMax has caught up with the now-identified "Space Bunny".
On October 17, DeepSeek — jokingly nicknamed "Big Fat Fish" by netizens — updated to V4.1 Flash-1017; just two days later, MiniMax officially released M3.1 Flash and claimed Space Bunny, the anonymous model that had drawn considerable attention in the developer community. With the two domestic Flash models appearing back to back, netizens have already started joking about "Big Fat Fish" and "Space Bunny".
However, this time MiniMax isn't really here just to compete on price.
According to the model evaluation data released this time,M3.1 Flash ranks first in the Automation-Bench comparison of complex Agent tasks, ties with Opus 5 for first on BankerToolBench, and comes in second only to DeepSeek in Coding cost-effectiveness. The two domestic Flash models updated back to back — one maintains its lead in software engineering, while the other delivered better results in complex Agent tasks.
On the Automation-Bench dataset, M3.1 Flash offers the best task-cost distribution cost-effectiveness
We also ran a practical task on M3.1 Flash: giving the model a hand-drawn floor plan and a building dimension table, and having it build from scratch an interactive 3D campus usable in a browser.
In terms of results, both solutions fulfilled the main deliverable requirements, but MiniMax Code + M3.1 Flash took 55.4 minutes, while OpenCode + DeepSeek V4.1 Flash spentabout 121 minutes. Converted at the API pricing used in this test, model costs on both sides were around 6.5 yuan.
With model costs of just a few yuan on both sides, MiniMax saved about 65 minutes.
01
Space Bunny Reveals Its Identity,
Where Does MiniMax Excel This Time?
Before officially revealing its identity, Space Bunny had already made a name for itself in the developer community.
OpenRouter disclosed that it had held the top spot on the platform's model usage ranking for 10 consecutive days, cumulatively processing 56.7T tokens. MiniMax also confirmed this time that the anonymous model processed more than 200T tokens cumulatively during relevant testing and usage.Fast speed,was the impression many developers had of it.
Developer Hassan also posted a very直观 intuitive comparison on X: he had M3.1 Flash and Claude Opus 5.5 each build a launch landing page for a hardware product; MiniMax's delivered design left him quite satisfied, and the bills on both sides were only 0.014 USD and 1.22 USD respectively. In his test, MiniMax also completed the task about 4 times faster.
Of course, a single developer test cannot represent a model's overall capability. But what's also noteworthy about MiniMax this time are the leaderboard results delivered after its official release.
As of October 20, on the Artificial Analysis (AA) leaderboard, M3.1 Flash's intelligence index reached 43 points, up from 29 points for the previous-generation M3 — an improvement of nearly fifty percent. Meanwhile, its task-weighted average model cost dropped from 0.51 USD to 0.22 USD.In other words, the score went up, while the cost of completing tasks fell by more than half.
M3.1 Flash scored 43 points on the AA leaderboard, ranking among the world's leading models
However, if it truly goes head-to-head with DeepSeek V4.1 Flash-1017, updated just two days ago, the two models each have their wins and losses.
On the software engineering benchmark DeepSWE v1.1, DeepSeek achieved 77.0%, while M3.1 Flash scored 73.7%. Although a gap remains, the latter is already in a range close to Opus 5 (74.0%) and GPT-5.6 Sol (73.0%). According to the coding cost-effectiveness comparison released this time, DeepSeek ranks first, with MiniMax close behind.
But on the complex Agent tasks that MiniMax focused on this time, the situation is reversed. In the Automation-Bench v1.0.6 model comparison released this time, M3.1 Flash surpassed DeepSeek and took first place; on BankerToolBench it scored 76.2%, tied for first with Opus 5. In addition, it achieved 82.2% on the cybersecurity benchmark CyberGym.
Image: benchmark leaderboard data
The content tested by Automation-Bench is also quite close to real work. It was built by the SaaS automation platform Zapier and contains 600 real enterprise business process tasks, covering scenarios such as sales, marketing, operations, customer service, finance, and HR. Rather than simply generating a piece of code, these tasks test whether a model can continuously call tools across different systems and carry a task through from start to finish. BankerToolBench, meanwhile, shifts the scenarios to financial business, placing higher demands on multi-step planning, cross-document information processing, and tool calling.
Seen this way, DeepSeek and MiniMax do not need to rush to decide an overall champion.On coding, DeepSeek still has the advantage; on complex Agent tasks, MiniMax has won one back. For developers, the domestic Flash option now has one more capable choice.
More importantly, beyond the performance gains this time, M3.1 Flash costs 1 yuan per million tokens for input, 4 yuan for output, and only 0.1 yuan for cached reads. Compared with the previous generation M3, input and output prices are roughly cut in half, and the cached-read price has dropped by about 75%; compared with DeepSeek V4.1 Flash-1017, its input and output prices are half of the latter's peak-period prices, on par with its off-peak prices, and it makes no peak/off-peak distinction throughout the day.
After the prices were announced, people in the comments were already making jokes about DeepSeek: "MiniMax is all about cost-effectiveness—with prices like these, who needs the big fat fish, right?"
One longtime user put it even more bluntly: "Not saying it's super powerful, but after falling behind for so long, it has finally caught up with the times."
Jokes aside, these two voices fairly well illustrate MiniMax's position this time. In the past, when outsiders talked about its text models, price was more often the topic than capability, but M3.1 Flash this time at least has a report card that can be compared head-to-head with mainstream models.Low prices are no longer the only selling point, and MiniMax finally has the chance to have developers discuss both its capability and its cost-effectiveness.
02
Hands-on test: building the same 3D campus,
MiniMax saved 65 minutes
The two Flash models on the leaderboard each had their wins, so we also assigned them the same practical task: based on a hand-drawn campus floor plan and a building dimension table, build a 3D campus from scratch that can be interacted with in a browser.
The materials provided by the task were a hand-drawn campus floor plan and a building dimension table, requiring completion of campus layout and conflict checking, writing a Blender script for modeling, keeping an editable source file and exporting GLB, then building an interactive webpage with Three.js supporting building selection, detail viewing, camera positioning, and simulated time.
The tests respectively used MiniMax Code＋M3.1 Flash and OpenCode＋DeepSeek V4.1 Flashrunning in the same local environment.
Campus hand-drawn floor plan and building dimension data table
▎Both turned in their work; MiniMax has richer interactive details
Looking at the delivered results first, both solutions completed the main requirements, but the pages were presented differently.
DeepSeek leaned more toward building data display, organizing information through a left-side list, a right-side detail panel, and a bottom time slider; MiniMax placed building numbers, site information, and view switching in more prominent positions, and after clicking a building, an orange highlighted outline and floating labels work together, making the correspondence between the scene and building data more intuitive.
Both kept editable source files, and the exported GLB models passed verification of six building numbers. For this task, both solutions delivered projects that could run and be further modified later.
▎At roughly the same 6.5 yuan, MiniMax turned in its work an hour earlier
What really made the difference was time.
According to local session statistics, the MiniMax combination took 55.4 minutes, while the DeepSeek combination spent about 121 minutes. The former used about 65 minutes less, with session duration cut by 54%.
The model bills on both sides were nearly identical. Converted at the official API prices verified on October 20, MiniMax's model cost for this task was about 6.50 yuan, while DeepSeek's, calculated at peak pricing, was about 6.52 yuan. MiniMax's cache hit rate in this run reached 98.1%, which also helped it keep the cost of multiple consecutive calls relatively low.
With the main deliverable requirements all completed and model costs close, MiniMax completed in less than an hour the work that took the DeepSeek combination about two hours. For Agents that need to continuously call tools and repeatedly modify code, waiting an hour less is itself a very tangible advantage.
03
Flash is starting to steal flagship models' business,
the price war now extends to per-task cost
In the past, Flash was mostly seen as a cheap substitute for flagship models.
Only when budgets were limited and tasks simple would developers consider Flash first; when facing complex software engineering or multi-step tasks, they still habitually handed work to more capable flagships. But as DeepSeek and MiniMax updated in succession, this choice began to change.
Especially in Agent scenarios, models need to repeatedly read files, call tools, modify code, and handle errors. A single task may last tens of minutes or even hours, and what developers care about ishow much time and cost the entire job needs to be completed.
The ALE cost comparison MiniMax published this time provides an intuitive example.
Among the 152 tasks in the full ALE-v1 Overall set, M3.1 Flash had a pass rate of 29.6%, with an estimated model cost of about 739 yuan for the whole benchmark; GPT-5.6 Sol, with a similar pass rate (about 30.6%), correspondingly cost around 5000 yuan.
ALE cost chart
The two differ by only about 1 percentage point in completion rate on this benchmark, yet their model costs differ by nearly 7 times.
Of course, the 29.6% pass rate also shows that M3.1 Flash is still far from being able to comprehensively replace flagships; in real use, failure retries and manual checks still need to be counted into the cost. But for some complex tasks, developers already have reason to try Flash first rather than immediately calling the most expensive model.
MiniMax will also later release a Pro version pursuing higher intelligence ceilings. And Flash's impact on flagships is already emerging:work that used to require calling a flagship can now, in part, be handed to the cheaper Flash. (Lei Feng Wang (Official Account: 雷峰网))
This is an original article by Lei Feng Wang; unauthorized reproduction is prohibited. For details seeReprint Notice。