Three frontier models in twelve days. February closes on a crowded market
Grok 3 on the 17th, Claude 3.7 on the 24th, GPT-4.5 yesterday. When releases arrive this fast, the buying question changes shape.
xAI released Grok 3 on 17 February. Anthropic released Claude 3.7 Sonnet on the 24th. OpenAI announced GPT-4.5 as a research preview on the 27th, describing it as a step forward in scaling pre-training rather than a reasoning model.
Three frontier releases from three labs inside twelve days. A year ago that would have been a quarter’s worth of news.
What this density actually tells you
No one is decisively ahead. When three labs ship within two weeks and the comparisons are argued rather than obvious, the honest reading is that top-tier models are converging in general capability. The differences that remain are in specific tasks, price, speed and terms.
Two strategies are running in parallel. GPT-4.5 leans on bigger pre-training; Claude 3.7 leans on variable thinking at answer time. Both were released the same week, which is a reasonable summary of a field that does not yet know which lever matters more.
Announcement fatigue is now a real risk for buyers. If you evaluate every release, you will evaluate forever and deploy nothing.
The buying question that survives this
Not « which model is best? » That question has an answer that expires in eleven days, as February has just demonstrated.
Ask instead: does the tool I am buying let me change the model behind it without rebuilding my process? If yes, the release calendar becomes somebody else’s problem and you upgrade when it suits you. If no, every one of these announcements is a small decision you now own.
Closing February
This month produced the AI Act’s first obligations, Switzerland’s decision not to write its own AI law, a summit that changed its own name from safety to action, the withdrawal of the EU’s liability proposal, and three frontier models.
Almost none of it changes what you should do in March. Pick one repetitive process. Clean the data it depends on. Decide who checks the output. Then run it for a month before starting the second.
That advice has survived every month of this blog so far, which is the strongest argument for it.