MiniMax Built MORE, DeepSeek Missed LESS: Same-Price AI Models TESTED

This is the video written out, with every figure in full, each linked to its source in the table below.

Two Flash models, two days apart▶ 0:00

On Saturday 17 October, DeepSeek released a new version of its cheap model, DeepSeek V4.1 Flash 1017. Two days later, on Monday 19 October, MiniMax released its own, MiniMax M3.1 Flash. They cost the same per token, as long as DeepSeek is on its off-peak rate.

We had both of them build things: a pagoda at dusk, a platformer game that plays itself, a solar system, a row of gears, a planetary gearbox, a Newton's cradle, and a picture of where the Sun and the Moon were at midnight UTC on 20 October. Where a test had a right answer, DeepSeek got it right every time. MiniMax missed in two of its twelve runs, and in one of them it worked out where the Moon was, wrote the right number on the label, and drew the Moon somewhere else.

Many of the numbers on this page are other people's: Artificial Analysis's independent tests, each company's own benchmarks, LMArena's voting board and posts on X. Each says whose it is. Our own tests are small, a few runs per prompt at default settings, and the page says so wherever they come up.

The same rate card▶ 0:41

Both models charge $0.15 per million input tokens and $0.60 per million output tokens. For DeepSeek, that is the off-peak rate. From 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, except on Chinese public holidays, DeepSeek charges double: $0.30 in and $1.20 out. That is 7 hours every weekday. MiniMax charges the same at all hours.

Only cached input is priced differently. That is the part of a prompt the provider has already seen and kept, which an agent re-reading a long conversation sends over and over. MiniMax charges $0.015 per million cached tokens, and DeepSeek charges $0.003 off-peak, a fifth as much ($0.006 at peak).

What M3.1 Flash is▶ 1:04

MiniMax M3.1 Flash came out on 19 October. It reads text, images and video, and its context window is 1M tokens. MiniMax gives its size as about 428B in total, 23B active per token. MiniMax says a Pro model will follow, and that the anonymous model called Space Bunny, which processed more than 200 trillion tokens, was an early version of M3.1 Flash.

The weights came out four days later, on Hugging Face. The repository was last changed at 03:37 UTC on 23 October, the date MiniMax had given in its launch post on X (“Oct 23”). They are open weights under the MiniMax Community License. That licence lets anyone use the model free of charge for non-commercial purposes and puts conditions on commercial use: a product built on it has to display “Built with MiniMax M3.1-Flash”, and its maker has to send MiniMax a one-time notice, or get MiniMax's prior written authorisation if the product earns more than 20 million US dollars a year. The licence also lists uses it prohibits. Because of the non-commercial grant, we call these open weights, not open source.

Against M3, MiniMax's previous model, M3.1 Flash costs half as much per token: MiniMax's price page has M3 at $0.30 in and $1.20 out per million tokens (for prompts up to 512K input tokens), against $0.15 and $0.60. On the Artificial Analysis Intelligence Index, an independent score built from 10 tests, it goes from 29.2 for M3 to 42.7.

MiniMax's release post says M3.1 Flash makes “online inference nearly three times faster and reducing usage costs by 70%”, that for agentic tasks its overall cost is 70% lower than M3's, and that its output speed in everyday use can reach 200+ TPS, with a median (P50) of 170+ TPS. Those are MiniMax's own claims, not measurements. Artificial Analysis's figures, read 25 October, don't show a threefold difference in speed: its median output speed was 95.6 tokens a second for M3.1 Flash and 88.7 for M3, both rolling medians. Its cost per index task was $0.220 for M3.1 Flash against M3's $0.508, 57% lower (our arithmetic).

DeepSeek's new model is an update to V4.1 Flash, which DeepSeek released on 10 September. Its weights came out the same day as the model, on Hugging Face, under the MIT licence. DeepSeek's own API serves it as deepseek-flash, with a context of 1M tokens.

Artificial Analysis, head to head▶ 1:51

On Artificial Analysis's Intelligence Index, DeepSeek comes out a little ahead: 45.1 against 42.7 for MiniMax, read on 25 October. MiniMax's score was 42.4 when we first read the page on release day.

Artificial Analysis also counts what each model cost to run its tests: $0.262 a task for DeepSeek and $0.220 for MiniMax. But it priced DeepSeek at $0.30 in and $1.20 out per million tokens, which is DeepSeek's peak rate. DeepSeek's price page says its off-peak rates are half of its peak rates, so the same run off-peak would have cost about $0.131 a task (our arithmetic).

Off-peak, most of the difference that remains comes from cached input. Of MiniMax's $0.220, $0.173 went on input, $0.121 of it on cache reads, and $0.047 on output. Off-peak, DeepSeek's input comes to $0.077, with $0.028 of cache reads, and its output to $0.054.

On that run MiniMax wrote fewer tokens than DeepSeek: 183M output tokens against 211M, most of them reasoning in both cases (151M and 171M). DeepSeek also read more input, 10.3B tokens against 8.0B. Artificial Analysis lists DeepSeek's row as “DeepSeek V4.1 Flash 1017 (Max)”.

MiniMax's own table▶ 2:27

MiniMax published a table of benchmark results with its release post on 19 October, and its Hugging Face model card of 23 October carries the same table with one cell changed. This section follows the card. It includes DeepSeek V4.1 Flash 1017. The table has fifteen rows, and DeepSeek has a score on six of them. On four of those six, DeepSeek is ahead, including the two coding tests that have a DeepSeek score, DeepSWE and NL2Repo. MiniMax leads on two: Automation-Bench, a test of AI agents on business workflows, 62.8 against 60.5, and Humanity's Last Exam, HLE, on a row that compares two different sets of questions, as the last paragraph here explains. MiniMax's scores for M3.1 Flash are its own runs, except GDP.pdf, taken from Artificial Analysis, and Agents' Last Exam, from that test's official leaderboard. The six DeepSeek figures are the same as the ones on DeepSeek's own model card.

Terminal-Bench 4.0, a test of tasks done at a command line, is in both companies' own tables and in Artificial Analysis's. MiniMax reports 38.1 from its own run in the Codex harness, and Artificial Analysis measured 25.8. DeepSeek's model card reports 42.3 from a run in Claude Code, and Artificial Analysis measured 33.8. Each company ran the test with its own setup, and both came out higher than the independent run.

The other nine rows have no DeepSeek score. MiniMax's note on its Agents' Last Exam row says the table takes only results from the official leaderboard and that a blank is not a zero. DeepSeek's own card reports 35.4 on that test. The GDP.pdf row, MiniMax says, is taken directly from Artificial Analysis, whose page also has DeepSeek 1017 on that test, at 17.2%, both when we read it on 19 October and on 25 October. That cell is blank in MiniMax's table, and MiniMax doesn't say why.

The HLE row sets two different things side by side. MiniMax's 43.4 is on the text-only questions (2,158 of them). The 42.6 beside it is DeepSeek's score on all the questions. DeepSeek's own card gives both: 42.6 on all questions and 44.0 on the text-only ones, which is just ahead of MiniMax. MiniMax's release post had the 44.0 in that cell, and its card changed it. Artificial Analysis ran both models on the same 2,158 text-only questions, and on that run MiniMax leads, 48.0% to 43.3%.

When it doesn't know▶ 3:28

Artificial Analysis also runs a knowledge test, AA-Omniscience, which in its own words “assigns points for correct answers, subtracts points for hallucinated responses, and keeps abstentions neutral”. MiniMax knows more: it gets 56.2% of the questions right, and DeepSeek 42.9%.

But when MiniMax doesn't have the right answer, it usually guesses: of the questions it doesn't get right, it answers 92.8% wrong. Artificial Analysis calls that the hallucination rate. DeepSeek's is 66.8%. Across all the questions, both give a wrong answer about four times in ten (40.6% for MiniMax, 38.1% for DeepSeek, our arithmetic). MiniMax just almost never holds back an answer: it leaves 3.2% of the questions unanswered or only partly answered, against DeepSeek's 19.0%.

On the test's own score, right answers minus wrong ones, MiniMax came out ahead in our 25 October read, +15.6 against +4.7, because it knows so much more. Artificial Analysis has since moved its index to version 5, and on it DeepSeek leads this score, 0.9 to MiniMax's -6.4 (read 7 November).

Potion, our podcast tool▶ 4:08

Potion is our podcast tool. If you watch this channel, you probably follow a few topics more closely than any podcast does. With Potion, you ask your AI, like ChatGPT or Claude, for a podcast episode on exactly what you want to understand. Potion narrates it and puts it in your podcast app, so you can listen on the go.

It works with ChatGPT and Claude, even the free versions, and it's free to start. Try Potion if you'd like your own personalized podcast feed, managed by your AI.

Our test: a pagoda at dusk▶ 4:52

We ran two kinds of test. In the first, there's no right answer: three prompts, each asking for a single web page, run twice per model on 19 October, at default settings, one shot, with nothing fixed by hand. MiniMax M3.1 Flash ran on OpenCode Go's free endpoint for it. DeepSeek ran through OpenRouter, pinned to DeepSeek V4.1 Flash 1017, and OpenRouter sent each run to one of three paid providers: InferenceNet (the two pagodas), Relace (the two platformers and one solar system) and Wafer (the other solar system). As of 25 October OpenRouter listed their quantization as fp8, unknown and mxfp4. We have no listing from the day of the runs, and we don't know how much of what follows that affected. On 25 October OpenRouter listed 12 paid endpoints for M3.1 Flash and none for its own free variant. Ours was OpenCode Go's free endpoint, which is a different route.

Each prompt is a list of things to include, so we checked every page against its own list, from a recording of the page (14 s with no key pressed and no click) and from its code. The first prompt:

Create a single self-contained HTML file that renders a 3D Japanese pagoda in a garden at dusk with three.js (load it from a CDN). The camera slowly orbits the pagoda; add lanterns that glow, falling cherry blossom petals and a small pond with reflections. Output only the HTML.

MiniMax's second page has everything on the list, and a lot that isn't: a title, Twilight Garden, a red arched bridge, koi in the pond, a full moon and stars. Its sky looks more like night than dusk, and the bamboo behind the pagoda sticks up above its roof. MiniMax's first page never got past its loading screen. The code declares a variable inside one function and uses it outside, so the page stops with “dummy is not defined” before it draws anything.

DeepSeek's two pages both run, with smaller slips. In the first, the falling petals come out as dark flecks against the bright sky. In the second, the pond's water only comes into the corner of the frame after about eight seconds.

The prompts also said to output only the HTML. All six of MiniMax's answers were only the HTML. All six of DeepSeek's added a sentence before the code and a section of notes after it.

A platformer and a solar system▶ 6:10

The second prompt:

Create a single self-contained HTML file with a small playable side-scrolling platformer game: a character that runs and jumps with the arrow keys, platforms, coins to collect, a score counter and one enemy that ends the game on contact. It should also play itself with a simple automatic demo mode when no key has been pressed for three seconds. Output only the HTML.

All four games have every item on the list, and all four demos started at 3.0 s exactly. MiniMax's two look finished, Neon Dash and Coin Rush, and both demos jump the enemy and keep collecting coins until the recording ends. The enemy never touched the player in MiniMax's recordings, so we checked in the code that contact ends the game. DeepSeek's two demos run into the enemy at about eight seconds (8.0–8.2 s and 8.4–8.6 s). It's game over, which is what the prompt asked the enemy to do, and the level starts again.

The third prompt:

Create a single self-contained HTML file with an animated, interactive solar system: the Sun and eight planets on orbits with relative speeds roughly right, a speed slider, and clicking a planet shows its name and one fact. Output only the HTML.

All four pages use the real orbital periods: every planet on every page is within 0.64% of NASA's figure. We clicked on Earth on each page. DeepSeek's pages show a small card with Earth's name and one fact. MiniMax's open full panels, with the orbital period and the distance from the Sun as well.

But MiniMax's first page draws the Sun and the planets off their own orbits, with stray arcs sweeping across the system, and its speed slider sits below the bottom of the window at the size we recorded. The code places each body by its corner, not its centre, and the trail arcs take the planet's angle as their rotation. MiniMax's second page has Earth's panel open before anyone clicks, so we opened Mars first, then clicked Earth.

Tokens and time▶ 7:25

MiniMax wrote more tokens on every prompt: 257,813 output tokens over its six runs, against 145,487 for DeepSeek, 1.77× as many. Most of them were reasoning, for both models. At the list prices, with DeepSeek off-peak, the six pages would have cost $0.155 on MiniMax and $0.087 on DeepSeek (our arithmetic; MiniMax's endpoint was free for us, so this isn't what we paid). On these prompts MiniMax wrote far more, even though on Artificial Analysis's run it wrote fewer. Three prompts at default settings show what these pages took, not a general rate.

MiniMax was also slower: 5.1 to 8.1 minutes a page, against 1.2 to 4.4 for DeepSeek. MiniMax ran on a free endpoint and DeepSeek on paid ones, so our times include the route.

Artificial Analysis's own speed test also has MiniMax slower. On 25 October its median output speed was 95.6 tokens a second for MiniMax and 195.2 for DeepSeek, and on its long prompt the first answer token, after the reasoning, came after 22.2 s for MiniMax and 11.4 s for DeepSeek. These are rolling medians that AA keeps re-measuring. We read 122.8 and 216.5 on 19 October, and 93.1 and 208.7 on 21 October.

Gears and a Newton's cradle▶ 8:08

The second kind of test has a right answer we can check. These ran later that night, just after midnight in Berlin on 20 October: four prompts, three runs per model, default settings, one shot, and both models through OpenCode Go. MiniMax used Go's free M3.1 Flash endpoint and DeepSeek used Go's deepseek-flash model. Go's responses don't say which upstream served them. When we last saw Go name it, on 17 October, deepseek-flash went to DeepSeek's own API, which serves DeepSeek V4.1 Flash 1017 under that name.

First, gears. One prompt asks for a row of five gears with 12, 30, 18, 40 and 24 teeth, the leftmost driven at 60 rpm. Another asks for a planetary gearbox: a sun gear with 24 teeth, three planets with 18, and a ring of 60 held still. Each page had to report its gears' positions and angles, so we read the speed and direction of every moving gear off the page, and we scored the teeth from still frames with the model's name hidden.

Both models got every speed and direction right, in every run. On the row, all five gears turned at the right speed and the right way. In the gearbox, the carrier turned at 17.14 rpm and the planets at −40 rpm, as they should. The teeth meshed cleanly every time.

Then a Newton's cradle. Pressing 1, 2 or 3 should lift that many balls on the left and let go, so the same number swings out on the right. DeepSeek's cradles got all 9 presses right. MiniMax's got 6. In one of its runs, pressing one does nothing, pressing two sends out one ball, and pressing three sends out two. The swinging is fine. The page takes one away from the key you pressed before it lifts the balls: launch(Number(event.key) - 1);

Where the Moon was▶ 9:10

The last test is the hardest. We gave both models one moment, midnight UTC on 20 October, and asked for a single picture, worked out without tools: a circle for the ecliptic, the path the Sun traces across the sky over a year, with the Sun and the Moon on it where they really were at that moment, and the Moon's phase in the middle. Then we checked each picture against JPL's DE440s ephemeris. The Sun was at 206.6° and the Moon at 311.2° of ecliptic longitude, with 0.627 of the Moon's disc lit, and growing.

DeepSeek's pictures are plain, and its Moon sits in the right place all three times. MiniMax's have more on them, craters and labels, and twice its Moon is in the right place too. The other time, it worked out the Moon's position, wrote the right number on the label, 311.28°, and then drew the Moon at 48.7°, 97.5° away. That's close to the mirror image of the right spot across the horizontal axis (48.8°). Both models put the Sun in the right place and got the phase right in every run.

DeepSeek took 2.4 to 3.1 minutes per picture. MiniMax took 7.3 to 19.0, on its free endpoint.

So the same pattern ran through most of it. MiniMax's pages mostly have more on them, and on our routes they took longer to write. DeepSeek's are plainer, and where there was a right answer, DeepSeek didn't miss. Each of MiniMax's two misses came one step after the hard part: the physics and the astronomy were right.

Other people's head-to-heads▶ 10:32

In the first two days, lots of people posted their own tests on X, almost all of them visual builds, and many of them against DeepSeek's new model. These are their own tests, with their own prompts, harnesses and prices, not a measurement.

We went through the 62 posts our searches of X found on 21 October. Eight of them compare MiniMax M3.1 Flash with DeepSeek V4.1 Flash 1017 by name and give a cost for each, and in six of those eight, MiniMax's run cost less. Two more gave both costs without saying which DeepSeek build they ran, one of them relaying another author's test, so we left them out. The three most viewed all had MiniMax cheaper, two of them at less than half the price:

• @Bhavani_00007 built a whale breaching, using Claude Code as the harness. MiniMax's cost $0.25 and DeepSeek's $0.76 (the DeepSeek cost is in the post's video). The author liked MiniMax's result best: “the result looks better than the others”.
• @rubenssoto_ai built a campfire: $0.22 for MiniMax, $0.54 for DeepSeek.
• @atomic_chat_hq, the account of an app for running AI models, rebuilt a police chase: $2.34 for MiniMax, $3.07 for DeepSeek.

The posters didn't agree on which one looked better, though. @EnvolDev ran an alpine flight with four budget models and put DeepSeek first: “DeepSeek V4.1 Flash 1017 blew me away”, with “MiniMax is right behind, and even cheaper” ($0.23 against $0.49, from the post's video). @RafaCrackYT had four models drive a roadster, where DeepSeek cost $0.66 and MiniMax about $1.37, and called MiniMax “the worst visually”. The other three: @EnvolDev's Eiffel Tower, @fawnunwrap's deer and @RafaCrackYT's rocket launch.

On LMArena's WebDev board, where people vote between two models' answers to web development tasks, MiniMax M3.1 Flash sat level with Claude Haiku 5.5 (High) and well below DeepSeek V4.1 Flash 1017 (Max) on the board dated 22 October, as we read it on 25 October: rank 34 with 1589, marked preliminary, against Haiku's 1586 at rank 35 and DeepSeek's 1665 at rank 14.

And @notjazii, after a few hours with M3.1 Flash on release day, found that it “hallucinates alot on xhigh or max and starts doing stuff i didn't ask for”. Those are its two highest effort settings. The same post calls it “a huge step up from m3” and says “in this test it did better than deepseek on first try”.

Which one, for what▶ 11:24

For coding agents and command-line work, DeepSeek: it is ahead on Artificial Analysis's independent run, and on the self-reported figures in MiniMax's table. Where a confident wrong answer costs you, DeepSeek too, narrowly: it holds back more often, though both give a wrong answer about four times in ten on Artificial Analysis's knowledge test.

MiniMax is worth a try for business automation agents, going by MiniMax's own Automation-Bench result, and for knowledge questions where you'll check the answer. For one-shot pages and games, MiniMax mostly made the fuller ones in our tests, but it wrote far more tokens for them, and in our tests with a right answer, DeepSeek was the one that didn't slip.

Per token, they cost the same with DeepSeek off-peak, except cached input, which DeepSeek sells for a fifth of MiniMax's price. On Artificial Analysis's run, that makes DeepSeek off-peak cheaper per task ($0.131 against $0.220), though in the head-to-heads people posted, MiniMax's run usually cost less. Both sets of weights are public if you want to run one yourself, DeepSeek's under the MIT licence and MiniMax's under its community licence.

Neither model is better overall. Which one to use depends on the work.