← The Polylog AI Intelligence Brief
Morning Edition · Tuesday, October 20, 2026Published at 2:12 AM EDT · New York
MiniMax prices its 428-billion-parameter M3.1 Flash at $0.15 per million input tokens and says weights ship October 23
The lab confirmed that the anonymous "Space Bunny" model, which processed more than 200 trillion tokens, was an early version of M3.1 Flash, and it reports 73.7% on DeepSWE v1.1.
Chinese lab MiniMax released M3.1 Flash on October 19 and confirmed that Space Bunny, an anonymous model that processed more than 200 trillion tokens during its stealth period, was an early version of the release, Unite.AI reported. The model is a 428-billion-parameter mixture-of-experts that takes text, image and video input with a context window of up to 1 million tokens, according to MarkTechPost.
The price is the central fact here. MiniMax charges $0.15 per million input tokens and $0.60 per million output tokens, and cached input costs $0.015. The company says online inference runs nearly three times faster than M3 and that usage costs fall by 70% (SiliconANGLE). The Russian-language channel AI ML Big Data reported that the weights will be released on October 23.
MiniMax reports 73.7% on DeepSWE v1.1, a benchmark of multi-step software tasks that requires the model to implement features, fix issues and validate its work with tests. The company also reports 48% on Humanity's Last Exam, ahead of DeepSeek and GLM, The Decoder reported. The Telegram summary adds 62% on Automation-Bench. All of these are vendor numbers, and none has been reproduced independently yet.
VentureBeat's assessment includes both a strength and a weakness. M3.1 Flash beats DeepSeek and Anthropic's Opus 5 on a business benchmark at lower cost, but it shows a high hallucination rate. That pattern matters for agent workloads, where a confident wrong tool argument costs more than a slow correct one. The 200-trillion-token figure measures how much the model was used during the stealth period. It does not measure accuracy.
For scale, Meta's Muse Spark 1.1 API, which The Decoder described in July as undercutting OpenAI and Anthropic, charges $1.25 per million input tokens. M3.1 Flash's list input price is one-eighth of that.
What this means
Open weights at 428 billion parameters move the price competition from MiniMax's own API to every inference host that can serve the model. That puts a ceiling on what US labs can charge for mid-tier and flash-tier models used in high-volume agent work. US closed-model vendors are exposed through their small and mid-size tiers, where margins depend on volume. Inference hosts gain a strong model to sell on serving efficiency. The constraint is the one described in today's DHS order: buyers in US government and defense-adjacent supply chains cannot adopt Chinese weights, so the price pressure lands hardest outside that perimeter. The consensus view treats each Chinese price cut as further evidence that prices are falling without any floor. The more specific effect is that the reference price for agent tokens is being set by downloadable models, and closed vendors must justify any premium through reliability (hallucination rates, tool-call accuracy) rather than benchmark scores. The likelier outcome is that within weeks of the weight release, third-party hosts in the US and Europe list M3.1 Flash at or near MiniMax's own price, and at least one US lab cuts or repackages its cheapest tier. This reading depends on the weights shipping on October 23 under a license that permits commercial use. A delay or a restrictive license would prove it wrong.
Part of a tracked trend
Frontier Model Price War
Frontier API vendors increasingly compete on strategic price-cutting rather than pure capability, repeatedly launching or repricing models below prevailing rates to grab share and compressing industry inference margins; expect recurring below-rival pricing moves.
Townsquare
Start a discussion in Townsquare.
More from this edition
- DHS orders civilian agencies to remove DeepSeek and any third-party product built on its models
- One hacker's AI agents breached more than 100 company websites at about $50 in tokens per target, Reuters reports
- OpenAI says GPT-6.1 Sol passes 86.6% of a new COBOL-migration test, up from 5.4% for GPT-5.6 Sol
- Germany and Nvidia sign letter of intent to fund industrial AI compute in Deutsche Telekom's Blackwell cloud
- Specialized agent guards block up to nine in ten safe actions when tool calls arrive as shell commands, ToolRosetta finds
- Prosecutors say accused OpenAI bugging suspect messaged a Saudi royal four days before his arrest
- PolyGraph captures one CUDA graph for adaptive speculative decoding by writing a 28-byte shape plan before each replay
Comments
0No comments yet.