AI labs focus on efficiency and reliability in new March 2026 models

In short: March 2026 brought a wave of new AI models that aim to make AI more useful by being cheaper to run, better at reasoning, and less likely to make things up.

What's going on

Big AI companies are releasing updates that focus less on size and more on day to day performance. OpenAI said its GPT-5 model cuts “hallucinations” (confident wrong answers) by 45% when it can use web search, and by 80% in a slower “extended thinking” mode. It also uses fewer “tokens” (small chunks of text that act like the model’s fuel) for tasks like reading images, writing code, and science problems.

OpenAI also described GPT-5.3 variants. One version, nicknamed “Garlic,” is designed to pack more knowledge into the same storage and it can handle a much larger “context window” (how much text it can keep in mind at once), up to 400,000 tokens. Another version, GPT-5.3 Instant, is aimed at smoother, more direct chat in ChatGPT.

Google’s Gemini 3.1 Pro also moved the focus toward reasoning and science knowledge, with large jumps on public tests that try to measure problem solving. Google also highlighted “Bayesian teaching,” a training approach that helps models update their beliefs as new evidence arrives (like changing your guess after each new clue).

What to watch

This shift could show up quickly in products people actually use. Faster and cheaper models make it easier to run AI on laptops and other devices, not just in big data centers. The same month also included more practical projects, like MIT work on predicting protein behavior to speed up drug research, and plans from Hyundai to put language models into robots.

Source: NYTimes