MiniMax Says M3.1 Flash Is 70% Cheaper Than Its Predecessor for Agentic Tasks
MiniMax says M3.1 Flash cuts agentic-task costs 70% from its predecessor and runs nearly 3x faster, claiming top-tier benchmark scores; weights are not yet released.
MiniMax has released M3.1 Flash, a model it says delivers a generational leap in agentic tasks, coding, computer use and cybersecurity while making online inference nearly three times faster and cutting overall usage cost 70% versus its predecessor. Output speeds can reach 200+ tokens per second, with median P50 above 170 TPS, which MiniMax attributes to more aggressive sparsity.
The company says the model ranks first among Chinese models on OSWorld V2 and Agents' Last Exam. MiniMax reports top-tier placement on AutomationBench-AA and 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82% on CyberGym, where it says internal security experts tested code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.
The bits that matter
- Space Bunny, an early version of M3.1 Flash, processed more than 200 trillion tokens while operating anonymously, according to MiniMax.
- MiniMax says M3.1 Flash can complete real enterprise workflows across SaaS systems in sales, marketing, operations, customer service, finance and HR.
- Flat pricing is $0.15 per million input tokens, $0.60 per million output and $0.015 per million cached input, with access through the MiniMax API, M Plan and MiniMax Code.
- Model weights will be released later, and MiniMax says a Pro model that pushes the boundaries of intelligence will also arrive later.
Discussion analysis
Analysis details
13 comments analyzed. Analyzed .
The sample selects active discussion branches and includes available parent comments. It may omit parts of the full thread. The analysis included 13 of 23 usable stored comments because of the input limit. Selected themes do not measure community opinion or how common a view is.
Cheap API models versus subscriptions
A self-reported workflow on DeepSeek Flash 1017 costs about $0.10 per project, and another report puts DeepSeek V4.1 Flash at roughly $1/day as a primary model. The counterargument is that Claude's $20 subscription only delivers a 60x API-value subsidy if the plan is maxed out through 5-hour and weekly limits; for work that cannot pause, API pricing is prohibitive. A separate objection disputes the subsidy framing and reports better value from a $40 OpenCode Go plan.
Throughput thresholds for agentic coding
A 13 TPS model is described as painfully slow for reading and unusable for agentic coding: if decode is that slow, prefill is abysmal, and agentic work depends on prefill speed. Chat does not need a 400B-parameter MoE model, but long tool-use sequences make throughput decisive. The claimed 200+ TPS and 170+ TPS P50 therefore bear directly on whether M3.1 Flash suits agentic workloads.
Benchmark baselines and model versioning
The graphics mix Opus 5 and Opus 5.5 reference points, described as odd and as mirroring a recent DeepSeek release. The choice is read as positioning against models that would not exist at current prices without competition. Because the benchmark scores are vendor-reported, the comparison's baseline and version consistency matter for interpreting the claimed parity.
Open-model lag and scaling costs
Open models are said to trail frontier releases by one to two months, partly because Chinese labs lack compute and must train longer or adapt to domestic hardware. A further concern is that pushing frontier performance may require 100x more compute for 2-3% benchmark gains, while 27-32B models could catch up on narrow coding tasks within a year.
Subscription AI and the feeling of being watched
One objection is that AI subscriptions require trusting a provider with conversations, and the sense of being recorded rules them out. The concern applies to hosted services generally, including MiniMax's API and M Plan, and lower token prices do not address it.
Duplicate submissions and HN policy
Five duplicate submissions of the same MiniMax item are flagged, with the note that submissions without discussion are not considered duplicates on Hacker News. The exchange is procedural rather than about the model.