A Brief Overview of This Week in AI
A quick look at a few things that happened in the AI space this week, including DoD negotiations, new free tiers, Cursor updates, and GPT 5.4.
2026-03-05
Loading content...
A quick look at a few things that happened in the AI space this week, including DoD negotiations, new free tiers, Cursor updates, and GPT 5.4.
This week saw Anthropic refuse to hand over its Opus model to the DoD on ethical grounds while OpenAI opted for an API-based military contract, highlighting the tension around defense deployments. In developer news, OpenAI dropped GPT 5.4 alongside offering Codex for free, and Cursor launched "Always on Agents" to automate your morning test suites.
The loudest fight this week was Anthropic and OpenAI negotiating with the Department of Defense. Same customer, opposite shapes for how the model gets used.
Anthropic drew a line it would not cross. The deal on the table was handing over Opus so the DoD could run it on government hardware. They keep the box. You keep arguing about deployment rules from a distance.
OpenAI took the other path. The DoD calls OpenAI's APIs. Every query routes through OpenAI's stack. Different risk profile, different visibility, different lever if something goes sideways.
I would not ship either arrangement on a product I care about without a lot more paperwork than a press release. The split still matters. Anthropic bet that giving up the model weights was the scarier concession. OpenAI bet that being the live pipe was acceptable. Both companies are selling trust. They picked different things to trade.
OpenAI shipped GPT 5.4. Another incremental bump on the chart, not a reason to rewrite your stack overnight.
I have not moved my default yet. I am running the same boring regression I always run. One gnarly refactor prompt, one long doc summary, same task on 5.3 and 5.4, compare the diff. If 5.4 wins on the work I actually do, I switch. Until then it is a bookmark, not a religion.
OpenAI put Codex on the free tier temporarily. If you have been curious about their coding models and did not want to pay to find out, open a tab this week. "Temporary" in vendor-speak usually means "until we have enough screenshots for the launch blog."
Cursor shipped Always on Agents. The pitch is morning chores without you in the chair. Pull branches, kick test suites, wait on builds while you are still making coffee.
When I was in big tech, the first hour was often git pull, CI, and figuring out what broke overnight. I would have pointed this at exactly that. Not because agents are magic. Because that hour was dead time I never got back.
I have not wired it into my current workflow yet. Solo projects do not have a standup waiting on green tests. If you do, try scheduling one routine Monday and see if you actually review output instead of rubber-stamping it.
I am watching whether the DoD deals change what "enterprise-ready" means in RFPs next quarter. On my machine, I am trying Codex on the free tier before it disappears.