GTC week: NemoClaw, GPT-5.4 mini and nano, and Mistral Small 4

Nvidia's GTC conference set the tone, and the tone was agents with seat belts.

NemoClaw

At GTC, Nvidia introduced NemoClaw, a stack that wraps OpenClaw-style personal agents in a sandbox with security and governance controls and managed inference. Jensen Huang praised the OpenClaw phenomenon on stage and then acknowledged the security problems it brought, which is why NemoClaw exists.

I find this the most telling item of the month. In late January the story was a viral agent project and the risk of giving it your accounts. Seven weeks on, one of the biggest companies in the world is selling the containment. That is the normal arc of a technology: first the demos, then the controls, then the budget line.

My own setup works the same way in miniature. Each agent gets narrow permissions, a log of what it did, and a human approval before anything leaves the building. Seeing the industry converge on that shape is reassuring.

Small GPT-5.4 models

OpenAI released GPT-5.4 mini and nano. Mini is meant for coding assistants and sub-agents, and nano is an API-only model for classification and data extraction at volume. Both cost more than the models they replace, which is a change from the usual pattern of each small model getting cheaper. I'll need to rerun my cost-per-task numbers before I move any workflow over.

Mistral

Mistral released Small 4, a mixture-of-experts model under Apache 2.0 that folds the company's reasoning, vision and coding lines into one. It also announced Forge, which helps companies build custom models on open foundations using their own data and policies. For firms that cannot send documents to a foreign API, European open models with a clear licence are an attractive middle path.

Looking at the price direction, I noticed the small GPT-5.4 models are more expensive than the ones they replace. After a year of falling prices, that is a change in direction, even if a small one. It might reflect higher capability, or it might reflect demand. Either way, it supports the habit of keeping a second model ready, preferably an open one I can host, so a price change does not hold my workflow hostage.

What I'd watch next: real pricing and quality comparisons of the GPT-5.4 small models against Qwen3.5 and Mistral Small 4 on high-volume extraction work.