Chinese lab MiniMax released its new flagship-tier agent model this week with a price designed to end arguments: M3.1 Flash runs $0.15 per million input tokens and $0.60 per million output — with cached input at $0.015, which the company says cuts overall agentic-task costs by 70% versus its predecessor while roughly tripling inference speed to 200+ tokens per second. Within a day it was topping agent benchmarks including AutomationBench, BankerToolBench (76.2) and LOCA-Bench-256K, hitting 73.7% on DeepSWE v1.1 and scoring 42 on the Artificial Analysis Intelligence Index — on par with GLM-5.3-Flash and 13 points above its own M3.
Then the launch turned into something more interesting than a price war.
The allegation
An AI researcher going by “Binkg” published a modified version of one of Artificial Analysis’ held-out AA-Omniscience questions — reworded so that the correct answer is different while sounding the same. Every model tested reasoned out the new answer. M3.1 Flash, the researcher claims, “just outputs the original Benchmarks answer. It doesn’t think about it but just recognizes the question in its thinking. Even on max effort.”
The theory behind the claim is plausible enough that Wccftech, which broke it down in detail, chose to publish it despite its colourful framing. AA-Omniscience is designed to be hard to cheat: only about 600 of roughly 6,000 questions are public. But MiniMax runs its own inference servers, which means it sees every prompt that reaches the model. Labs typically release API access before weights, so if an evaluator like Artificial Analysis runs a benchmark against a hosted endpoint, the host can in principle log those distinctive queries. Benchmarks published in June for M3 could then, in the worst case, have seeded M3.1 Flash’s training.
It is essential to state clearly what is actually established: this is an allegation by a researcher, echoed and amplified by one outlet, with no published confirmation. Wccftech says it contacted Artificial Analysis and will publish its response when it arrives; MiniMax has not addressed the claim as of publication, and this story has not been independently reported by a second outlet. Nothing in the public record proves query sniping occurred.
What the independent data does show
Separate from the accusation, independent evaluators’ numbers carry their own caution flags. According to Artificial Analysis’ own model page and analysis synthesised by i10x, M3.1 Flash scores 56% factual accuracy on AA-Omniscience — and when it answers incorrectly, it hallucinates 93% of the time, i.e. it is wrong with high confidence rather than hedging. For the enterprise workflows MiniMax is targeting (BankerToolBench is literally a financial-workflows benchmark), that reliability gap can consume the 70% savings on validation overhead alone.
There is also history here. We covered Berkeley’s finding that benchmark scores have become unreliable in April, and tracked AI coding agents gaming SWE-Bench Pro through task leakage in May. The MiniMax allegation is a sharper, allegedly mechanical version of the same failure: if a lab both serves and scores the model, held-out is only held-out until it isn’t.
Why the price signal still matters
Strip the controversy out and the launch still moves the market. MiniMax and DeepSeek have been trading benchmark-topping, bottom-dollar releases all year — and New Zealand buyers are downstream of that fight, whether via resellers or via local agencies building on Chinese API tiers. At $0.15/$0.60 per million tokens, M3.1 Flash undercuts the already-discounted DeepSeek V4.1-Flash tier we covered in our price-war analysis and pressures every Western lab charging ten to sixty times more for comparable agent throughput.
The uncomfortable synthesis: if the sniping claim is true, benchmark scores are purchasable, and buyers get burned the moment a model meets live traffic. If the claim is false, the model is genuinely cheap and capable with a documented hallucination problem — which burns buyers a different way. Either way, the lesson converges. Artificial Analysis evaluates models it does not host; that separation is the integrity model — but the MiniMax story shows even separation fails if the evaluated endpoint can log prompts. The 2026 fix is likely to be adversarial: rotating private question banks, live-probe spot checks, and buying on measured outcomes rather than launch-week leaderboards.
Sources: MiniMax announcement, Wccftech, Artificial Analysis, i10x analysis
