MiniMax M3.1 Flash: 42 on the AI index for $0.22 a task
China's open-model price war found a new floor today — and within hours, Meta showed what the other end of the market costs.
MiniMax released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index, 13 points above its predecessor M3, at less than half the cost per task. The score ties GLM-5.3-Flash, edges past Gemini 3.8 Flash at 41, and trails DeepSeek's V4.1-Flash-1017 at 45 — the model that held the open-weight lead for roughly two days before this launch, according to Artificial Analysis. The API is already live at platform.minimax.io; MiniMax says open weights land on October 23.
The interesting part is not that a Chinese lab shipped another competitive Flash model. It is that the axis of competition has moved from capability to cost per completed task. Artificial Analysis estimates M3.1 Flash finishes a task on about $0.22, against $0.26 for DeepSeek, because it emits far fewer output tokens across the same suite — roughly 189 million versus 224 million. MiniMax's own chart puts DeepSWE v1.1 at up to 73.7 percent for under $0.30 a task, and the company claims the model runs three times faster than M3.
Where M3.1 Flash wins, it wins on reasoning that requires actual knowledge rather than pattern completion. It scores 48 percent on Humanity's Last Exam, ahead of DeepSeek's 43. On CritPt, a research-level physics benchmark of 71 problems, it manages 22 percent — behind DeepSeek's 24, but clear of GLM-5.3-Flash at 15 and Gemini 3.8 Flash at 18. On AA-Omniscience, which measures factual accuracy, it answers 56 percent correctly against DeepSeek's 43 and GLM-5.3-Flash's 28.
The caveat is severe and worth stating plainly: Artificial Analysis records a 93 percent hallucination rate on that same factual test, because the model attempts every question rather than declining when it does not know. A model that is cheap, fast, and confidently wrong is a specific kind of hazard — fine inside an agent loop that verifies its own work, dangerous as a direct answer engine. That is the trade buyers are actually making at this price.
The launch also closed a loop the industry had been speculating about for weeks. "Space Bunny," an anonymous model that sat at the top of OpenRouter's rankings for ten consecutive days, was an early version of M3.1 Flash. OpenRouter says the stealth model processed 56.7 trillion tokens on its platform — the most of any anonymous model it has hosted — and M3.1 Flash is free there for 72 hours. MiniMax claims the earlier build handled more than 200 trillion tokens in total while unnamed. On the Code Arena WebDev leaderboard it debuts at 1,588 points, 12th among open models and 32nd overall, a 106-point jump over M3.
MiniMax is not done. Staff and Reuters-sourced reports point to an M3-Pro with roughly 2.7 trillion parameters in the pipeline, which would put a dense-class model against the frontier rather than the mid-tier. For now the practical read is simpler: the $0.15 input / $0.60 output pricing sets the reference point that every Western lab's "cheap tier" now has to beat, and the gap between the cheapest competent model and the most expensive one is widening, not closing. We looked at what that squeeze did to the previous record holder — DeepSeek V4.1-Flash-1017: 77% on DeepSWE for 20 cents — and the same logic applies here with a bigger context window.
Tenstorrent raised $1.5 billion at a $6.2 billion valuation, more than doubling the $2.6 billion it fetched in late 2024. The round was led by Trace Capital, the new firm from Quiet Capital co-founder Ben Mahdavi, and Innovation X, with Cadence Ventures, Eclipse, Fidelity, SanDisk and T. Rowe Price participating, according to Bloomberg. It lands five months after Intel and Qualcomm reportedly floated acquisition interest at up to $5 billion, and CEO Jim Keller told Bloomberg the company chose independence after demonstrating that its Galaxy servers could serve DeepSeek models.
The strategic detail is in what the money is for: chips tuned for serving open-weight models and video generation, not training. That is the correct bet. If MiniMax can sell frontier-adjacent intelligence at $0.15 per million input tokens, the margin migrates to whoever can run those weights most cheaply per token — and Nvidia's grip is weakest exactly there, where the workload is predictable and the software stack is less exotic. Tenstorrent's $6.2 billion is still small next to Etched, which TechCrunch reported was fielding offers above $40 billion this month. The gap between those two numbers is the market's guess about which inference architecture wins.
Meta published the engineering detail behind Prometheus, the gigawatt-scale supercluster in New Albany, Ohio, that trained its Muse models — and confirmed the scale: hundreds of thousands of GPUs today, roughly one million at full build-out. The first phase is complete and runs a mix of Nvidia GB200 NVL72 racks ("Catalina") and GB300 NVL72 racks ("Clemente"), with Vera Rubin NVL72 arriving via a platform called San Miguel, per Meta's engineering blog. Meta says the cluster sustained more than 85 percent goodput on pre-training runs and more than 90 percent on reinforcement-learning runs of Muse Wonder.
The number worth remembering is not the million GPUs. It is the interruption rate. Meta says early Prometheus hit nearly 1,000 interruptions per 8,000 GPUs per day in mid-2025, cut that by 99 percent over a year, and now targets two interruptions per 8,000 GPUs daily — about 100,000 GPU hours between failures, up from a measured 47,000 during the Muse runs. To make that usable it reserves eight of every 72 GPUs in a rack as in-rack spares and keeps about 2 percent of racks in reserve, so a 32,768-GPU training run provisions 523 racks and 37,656 GPUs, roughly 13 percent more hardware than the job uses.
That is the honest accounting of frontier training in 2026: most of the cost is idle silicon. The cluster also swings 30 to 40 percent in load within seconds when jobs start or pause, which Meta absorbs with power capping, rack batteries and capacitor trays. Next up is Hyperion in Richland Parish, Louisiana, designed with Prometheus' lessons baked in from the start.
What to watch: whether MiniMax actually ships weights on October 23, and whether Meta's interruption target holds as it adds Vera Rubin to a cluster that is already running two GPU generations at once.
Would you run a 93 percent hallucination rate in production if the tokens cost a fifth as much? Write to us via our contact page.
