Chinese AI startup MiniMax launched an updated version of its AI model M3.1 Flash on Monday, announcing the release in a statement reported by Reuters.

According to the company's own claims, the updated model delivers "generational" improvements in three areas that matter most to enterprise buyers right now: agentic tasks, programming and cybersecurity.

MiniMax also said the model's online inference speed has nearly tripled while usage costs have been reduced by 70%.

The combination being claimed here is unusual: nearly tripled inference speed and a 70% reduction in usage costs are the kind of efficiency gains that, if they hold up under independent benchmarking, would force competitors to respond on price.

It also lands just an hour after DeepSeek's own cheap, fast update grabbed the top open-model spot on coding leaderboards, showing Chinese labs are racing each other on the cost-performance frontier rather than only chasing Western flagship models.

MiniMax's focus on agentic tasks and cybersecurity also mirrors where Western rivals are heading, suggesting the competition is shifting from raw benchmarks toward practical, deployed workloads.

The details are so far the company's own: Reuters reports that the "generational" improvements in agentic tasks, programming and cybersecurity, the nearly tripled online inference speed and the 70% cost reduction all come from MiniMax's statement, with no independent benchmarks cited in the coverage.

The Standard in Hong Kong, covering the same announcement, notes MiniMax trades publicly under ticker 0100, meaning the market will be able to weigh these claims against reported revenue and usage directly.

In August the company reported that revenue nearly quadrupled in the first half as AI demand surged, providing context for how quickly its commercial base is growing.

A 70% reduction in usage costs, if accurate, changes the economics of running agentic systems, where a single user request can trigger dozens or hundreds of model calls.

Cheap, fast inference is the enabling layer for the autonomous-agent push that labs on both sides of the Pacific are now competing over, and it is also the pressure point in the current AI infrastructure selloff, with investors questioning spending across the stack.

For MiniMax, undercutting rivals on price while claiming gains in exactly the workloads enterprises care about — coding, security and agents — is a direct shot at both Chinese competitors like DeepSeek and Western providers charging far more for comparable capabilities.

The announcement lands in a crowded week for AI model releases, with Meta shipping open-weight coding and agent models and Google launching its Gemini 4 Argon flagship for developers, as The Standard's related coverage shows.

MiniMax listing publicly as 0100 and reporting surging first-half revenue makes its efficiency claims more than a research footnote: they are a commercial argument to investors and customers alike.

The key thing to watch is verification. MiniMax's speed and cost figures are self-reported in a company statement, and independent benchmark results on agentic tasks, programming and cybersecurity — the areas where it claims generational improvement — have not yet appeared in the coverage.

If the claims hold, expect another round of price pressure across the model market; if they do not, the gap between marketing language and measured performance will be visible fast.