Diffusion LLMs hit 1,068 tokens/s as MiniMax claims 70% cheaper agents

The cheapest way to make AI agents safer may turn out to be making them cheaper to run, and this week's news pulls those two threads closer together than the industry likes to admit. MiniMax says its new M3.1 Flash cuts the cost of running agents by 70%. I'd treat that figure with some scepticism until independent benchmarks appear, but the direction matters more than the exact number. When agent costs fall that sharply, people stop running one carefully supervised agent and start running dozens, and every one of them is another thing that can go wrong.

Speed is attacking the same problem from a different angle. Inception's Mercury 3 generates text through diffusion rather than the token-by-token process that conventional LLMs use, and it reaches 1,068 tokens per second. That is genuinely a different architecture, not an incremental tweak, and I find it more interesting than yet another transformer refresh. Whether diffusion models can match the reasoning quality of the best autoregressive ones remains an open question. But for agent workloads, where latency compounds across dozens of chained calls, raw speed can matter as much as raw intelligence.

Now consider how the infrastructure is catching up. Microsoft has made Foundry Routines and Tool Search generally available, so agents can run on a schedule and find the tools they need on demand. By the way, that second feature is easy to overlook, but an agent that discovers its own tools is an agent whose capabilities you can no longer fully enumerate in advance. This is exactly why the AWS security guidance circulating this week, with isolated execution environments, least-privilege IAM roles, tool-call validation and emergency kill switches, reads less like a niche how-to and more like a baseline requirement. If you are building with agents, deterministic boundaries around them are no longer optional.

The policy debate, meanwhile, is stuck at a much higher altitude. Andrew Ng has accused Dario Amodei, Sam Altman and Elon Musk of inventing safety arguments to justify doomsday framing, while Anthropic lobbies Congress for a federal bill that would let only the Commerce Department halt a model's release, and only if it risks $1 billion in damages or 50 deaths. I think Ng's critique has some bite, since regulation written by the most heavily capitalised players tends to favour the most heavily capitalised players. But a threshold like that also deserves scrutiny from the other side. Who measures a billion dollars of potential damage before a model ships? And Anthropic is simultaneously being sued by Postmedia, Torstar, The Globe and Mail and The Canadian Press in an Ontario court over how that content was used, which is a reminder that the near-term legal risks are far more mundane than extinction.

Perhaps the real safety question isn't whether a model could end the world, but who is accountable when thousands of cheap, fast, scheduled agents quietly make mistakes at scale.