DeepSeek 1st, MiniMax 2nd!

  It's the first time in the history of the world.

DeepSeek has just played, and another king has blown up!

Two days ago, DeepSeek released V4.1 Flash-1017, putting "Voice Programming" on the stage.

Just over a weekend later, the MiniMax was hit hard, and the M3.1 Flash had just been officially released!

The cost of coding.DeepSeek 1 and M3.1 Flash 2。

On the Agent mission, it's directly down.Automation-Bench first overallI'm going to run the ALE Agent Final.It's only $739.！

The price of the M3 was cut in half compared to the previous generation.Sexuality costs direct execution！

In just two days, two of the country's biggest models have joined forces, and the competition for the Flash segment is full!

On the official release list, M3.1 Flash and DeepSeek, GLM, GPT-5.6 Sol, Claude Opus 5 are the same as 15 models.Take three firsts out of five.；DeepSWE programming to align with the flag，Cybersecurity is on the front line.。

DeepSWE v1.1 is a software engineering benchmark that examines an entire development link, code understanding, function development, troubleshooting, test verification, all of which have to come from the model itself.M3.1 Flash scored 73.7 on the max.The GPT-5.6 Sol (73.0 points) and the GLM-5.3 (66.9 points) are pressed behind the OpenAI flag.

The M3.1 Flash cost a single mission on DeepSWE, and the M3.1 Flash cost a single mission on DeepSWE.Less than 2 yuan。

At this price, only the newly released DeepSeek V4.1 Flash-1017 is ranked higher.

The same goes for a score of 73% up and down, and Claude Opus 5 costs more than $70 for a task.It's 40 times faster than the M3.1 Flash.。

The two brothers, DeepSeek and MiniMax, are hard to come by.If you're going to get the flag-level score, you're going to get a penny or two.。

That's what makes the M3.1 Flash crazy.CoWork and Complex Agent workflows。

The Automation-Bench is built and open-source by Zapier, with 600 tasks all coming from real-world business processes, covering sales, marketing, operations, customer service, finance, and HR. It examines whether a model can work from start to finish across a set of SaaS systems.

M3.1 Flash is here.62.8All right, first place.！

It does.Open Opus 5 by 12 minutes.GPT-5.6 Sol was unlocked for more than 16 minutes, and even the recently released DeepSeek V4.1 Flash-1017 ((60.5 points) was unlocked.

More frightening is the cost. On this cost-and-score chart, all other models fall in the lower right corner of the M3.1 Flash, costing more and scoring lower.

M3.1 Flash only runs to medium files, and a task takes about 2 to 5 minutes to complete.The score has already been surpassed by the updated generation Opus 5.5 full of max files.Opus 5.5 runs a task that costs about $6.More than 20 times the price.。

Even at the highest level, a task costs around $3.00 to run 1,000 real-world enterprise tasks across SaaS systems.The bills are down to 300 yuan.！

M3.1 Flash on the LOCA-Bench 256KIt's also the first of its kind.It was released in the United States on September 5, 2003, and was released in the United States on September 6, 2007, and in the United States on September 7, 2007, and in the United States on December 7, 2007.

The BankerToolBench of the financial landscape examines multi-step planning, cross-document consistency, financial compliance and toolchain calls.It's the same as Claude Opus 5.GPT-5.6 Sol (74.6 points) is in the back.

The cybersecurity test CyberGym covers 1,507 historical vulnerabilities in 188 real software projects, with M3.1 Flash scoring 82.2 points, and GLM-5.3 with 84.5 points, GPT-5.6 Sol with 83.6 points.Standing on the world's first ladder。

In the field of computer use, the explosive power of M3.1 Flash is even more terrifying.They've taken the national model number one.Running the entire ALE.It's only $739.！

OSWorld 2.0 was launched by XLANG Lab headquartered in Port-au-Prince with 108 long-range missions covering seven major areas, with skilled people doing one task, typically taking about 1.6 hours.

M3.1 Flash, take it off.61.7 pointsThis is 7 points higher than the DeepSeek V4.1 Flash (54.3 points).

Agents' Last Exam (ALE), nicknamed "The Last Exam of Agents", is led by Berkeley RDI with hundreds of industry experts, covering 13 major industries and 55 segments.

M3.1 Flash's pass rate is29.6%He is also known for his work on GPT-5.6 Sol, 30.6%, and Claude Opus 5.32.2%.

A total of 152 ALE missions, or 739 yuan, is the estimated total cost of running the entire set.

In the cost chart of the ELA, in more than 20 models,It costs the least.！

The cheapest GPT-5.6 Luna is also around 1600 yuan, which is more than twice as much.

GPT-5.6 Sol scored just one point and spent over $5,000; Claude Opus 5 and GPT-6 Astra over $7,000; Claude Opus 5.5 was close to $9,000.

This is the first time I've seen this.The bills are over 10,000 yuan.The pass rate is only 21 percent and 16.5 percent, respectively, behind the M3.1 Flash.

M3.1 Flash is the starting point of this line, connecting the models with the highest scores at each price.

More than 700 to win 10,000The bills are dozens of times worse!

For every million tokens,Input is only 1 yuan, output is 4 yuanIt's the first time I've ever been on a mission, and I've been on a mission.The overall cost of use is 70% lower than its predecessor。

Similarly, the domestic version of Flash, DeepSeek V4.1 Flash-1017 is available in both high-end and low-end versions.

A lot of businesses are focused on day-to-day operations, and they hit the peak of DeepSeek, with 2 yuan in and 8 yuan out.The whole thing is twice as expensive.。

M3.1 Flash does not divide time.It's the price of DeepSeek free time all day.。

The M3.1 Flash's input cost was $190 at a 90% hit rate, compared to $236 at the peak of DeepSeek.It's $46.The hit rate is mentioned as 95%, both are 145 to 138 yuan, which is basically equal.

The ability is enough, the bills are low enough, and the Flash has a chance to take the lead.

Light is not cheap enough to keep up with long-range tasks.Daily output speeds of up to 200+ TPS，The median is stable at 170+ TPSThe last generation of the M3 is only 100+ TPS in official documentation.

The people who write the code know that kind of crap, the model repairs a round, finds there are problems, then waits for the next round.

This is precisely the strength of the M3.1 Flash. Its predecessor, the mysterious model "Space Bunny" that surrounds the entire network, is the fastest.

The official MiniMax statement, "Welcome back, Space Bunny", was immediately acknowledged.More than 200T Tokens！

The official version continues this very fast gene and replaces the more radical sparse attention structure.Online reasoning has nearly tripled in speed。

You just knocked on the door of the car, and the code on the other side of the car has already filled the screen.

The AI developer, Ivan Fioravanti, has successfully tested the API."Hurry up, hurry up".！

But it's more important than speed, this time it's the MiniMax.He fought a beautiful turnaround.。

For some time, the outside world was somewhat stereotyped about the MiniMax. When the last generation of the M3 was released, there was little controversy, and some even claimed that "the MiniMax is not good at making text models".

Today's M3.1 Flash, which we're using today's M3.1 Flash,In the meantime, I'm going to try to find a way to get back on track.。

The MiniMax itself is also built on it. In the list of models in the official API documentation, the M3.1 Flash has been ranked first.

M3.1 Flash native supportOne million tokens in the contextIt's easy to read text, images, and videos directly.

Thinking patterns areFrequently openI'll think twice before answering every question.

However, if you want to be fast or deep, you can give the effort parameters to adjust the five grades of low, medium, high, xhigh, max without filling in the default max.

Simple daily instructions can save you money and save you time. Complex system architecture rebuilds will fill up to max, and the model takes the problem one step at a time, returning the thought process separately, and delivering only the cleanest results in your hands.

Now, M3.1 Flash is officially available on the MiniMax API, M Plan and MiniMax Code.

Official CLI warehouse, the default text model has switched to MiniMax-M3.1-Flash

We incorporated M3.1 Flash into the codex and created a puzzle that would fold a city like a 3D book and make it into a playful 3D web page.

The blocks that stood on the ground were slowly rising, the roads became high walls, and the group of buildings followed the central fence.

Drag the sliders, and the city can stop at any folding point; cut through the night, a row of windows lights up, and the whole scene changes a face. Drag the picture, and you can go around the other side to see how the streets and buildings are connected.

A hole in the brain that tastes like a thief's dream space is a 3D toy that you can play with your own hands.

Developers AJ also used it for a series of tower-generating tests.It's one of the most visually impactful models he's ever measured.Even the DeepSeek V4.1 Flash is better than the previous one.

What's even more exaggerated is that the whole process was blown up!

There are no rounds of dialogue, no tools, no frameworks.A single API callThinking about filling, it cost about 50,000 tokens, and M3.1 Flash built the entire tower.

There are two flashes on the front and back of the table.The "Twin Stars" of the National Flash track have been formed.！

From this round of evaluation, DeepSeek has a stronger lead in DeepSWE software engineering tasks; MiniMax has a stronger lead in enterprise workflows and financial instrument calls for these tasks.

Both have home fields, which developers can use entirely, selecting models by task, and finally allowing success rates, time and bills to speak for themselves.

It's more like a real day-to-day use by developers than it is to get an overall list.

The biggest change is:Flash is reaching out to the work that needs to be flagged in the past.。

In the past, developers were willing to pay more to complete complex tasks in exchange for higher success rates and fewer folds. Now that Flash has caught up with the flagship in many tasks, the premium has to be recalculated.

Fixing a bug, running a business process, completing a series of computer operations, as long as cheaper models can deliver results, users have reason to try it first.

As DeepSeek and MiniMax continue to move forward, more and more live will be handed over to Flash, leaving the budget to the really difficult parts.

Pro's capabilities, Flash's speed and price.

While M3.1 Flash is just a snack, the official has spoken out in pursuit of the smart limit.The Pro version will come later.！

The national Flash Twins are on the table!