MiniMax releases M3.1 Flash for business automation
VentureBeat reports access through MiniMax Code and the M Plan subscription. Its testing coverage distinguishes completing business tasks from handling uncertain facts.
The research desk's take → New method improves text-model accuracy by up to 6.7 percentage points30-second briefing
- VentureBeat reports MiniMax releases M3.1 Flash for enterprise workflows.
- Availability includes an API, subscription, and desktop application.
- Watch factual-error handling and independent coding evaluations.
One-Line Summary
Low-cost business AI makes the gap between completing tasks and producing trustworthy answers central to choosing a model.
Industry & Biz
MiniMax releases a cheaper model for business-software tasks
MiniMax, a Chinese company that develops AI models, releasesM3.1 Flash to carry out tasks across business software, according to VentureBeat's October 19, 2026 report. The coverage describes applications in sales, marketing, customer support and finance, and cites MiniMax's claimed62.2 score on Automation-Bench, a test of business workflows. These are benchmark results, not evidence of a live company rollout. 1
As of October 26, 2026, the Artificial Analysis results reported byVentureBeat put the average evaluation-task cost at approximately**$0.22**, compared with**$0.51** for the preceding M3 model. For teams comparing options for repetitive work, cost per completed task is more informative than the advertised usage rate alone—but it does not include every expense involved in running and checking the system. 1
In the factual-knowledge testing covered by the report, M3.1 Flash answers 56% of questions correctly, alongside a93% hallucination rate under that test's definition. Crucially,93% does not mean that 93% of all answers are wrong: it measures how often non-correct responses are outright wrong rather than partial answers or refusals. For customer research or financial-document work, the concern is therefore not just making mistakes, but supplying an answer instead of acknowledging uncertainty. 1
VentureBeat lists usage prices of $0.15 per million input tokens and**$0.60 per million output tokens**, with tokens being the small text units used for billing. It reports access through theM Plan subscription andMiniMax Code desktop application, without specifying their subscription prices in the supplied coverage. For a non-technical team, the practical next step is to request current access, pricing and data-handling information—not to treat a benchmark score as approval for workplace use. 1
Community Pulse
Hacker News (517↑) — Commenters question benchmark relevance, citing weaker coding and cybersecurity results and doubts about value. 2
"Yeah, ollama has it up for free at the moment. I'm fairly nonplussed about it. Got it to do a code review on a couple of codebases where I knew roughly what the problems were and it underperformed even with a prod in the correct direction. Its code output is average, and it's more expensive than glm-5.3-flash, I don't see myself using it when it comes off free." — Hacker News 2
"I've been building a cybersecurity platform (called ZeroQuarry) and regularly benchmark new models against a pretty wide set of security both code review tasks and live application pen testing tasks, and this one far underperforms relative to GLM 5.3 from every test I've done for real-world cybersecurity tasks" — Hacker News 2
What This Means for You
For marketing and planning work, distinguish producing a useful draft from verifying the facts inside it. The reported business-task score and factual-accuracy results measure different abilities, so neither should substitute for checking whether a customer claim, figure or recommendation has supporting evidence. 1
For team purchasing discussions, compare the cost of an acceptable result rather than the usage fee alone. A simple evaluation sheet can record the output, corrections required and time spent checking it alongside any quoted price; this makes human review visible instead of treating it as free. 1
The Hacker News comments are individual accounts, not a controlled comparison that settles which model is best. Their practical value is the question they raise: does a published score match performance on the specific work your team needs? Ask for evidence from a comparable task before approving a purchase. 2
Action Items
- Audit an AI-written draft: Choose a non-confidential document you already have, highlight its factual claims and check each against the original material. Mark unsupported claims separately from writing or formatting problems.
- Draft a product-information request: Ask your IT or purchasing contact to obtain current M3.1 Flash access options, subscription prices and data-handling terms. Keep this as an information request rather than a request to upload company documents.
- Write a human-approval checklist: Choose a task such as customer-support replies and document what must be checked before an AI-generated answer is sent. Include factual support, missing information and who gives final approval.
Terms in This Article
- Hallucination
Hallucination is a failure mode where an LLM produces fluent content that conflicts with source evidence, real-world facts, or task constraints. It can come from gaps in training data, weak grounding, retrieval misses, or inference-time error propagation, so production systems manage it with source checks, uncertainty signals, and abstention policies.
- output tokens
Output tokens are pieces of text generated by an AI model in response to input, where the model predicts the next most likely tokens to produce a coherent and contextually relevant output.
- Inference
Inference is the execution phase where a trained model receives new inputs and computes predictions, classifications, or generated outputs. For LLMs, prompts are tokenized, processed through prefill and decode, and emitted as tokens while latency, throughput, and memory usage become core operating metrics.
Two desks on this story
MiniMax releases M3.1 Flash for business automation
New method improves text-model accuracy by up to 6.7 percentage points
T2Quant reallocates activation precision across denoising steps, reporting gains of up to 6.7 percentage points. A separate study shows why output length alone misses task- and language-dependent decoding costs.
Comments (0)