Skip to content
Free · 1,000+ readers
Free · Independent
The daily record of artificial intelligence
← Back
Frontier Models

Google DeepMind unveils Gemini Ultra-4, claiming first dependable long-horizon agent

In a closed demonstration to enterprise customers, the model completed a 47-step procurement workflow autonomously — a benchmark every laboratory has chased for eighteen months.

Saturday, May 16, 2026 · 4 min

Google DeepMind on Saturday introduced Gemini Ultra-4, a model the company describes as its first system capable of dependable multi-step autonomy in commercial workflows. The launch took place at an invitation-only event for two hundred enterprise customers.

In a closed demonstration, the model completed a forty-seven-step procurement workflow without human intervention, identifying suppliers, drafting contracts, negotiating terms and producing a final report. The same task, executives said, took its predecessor more than four hundred steps and required human correction at sixty-eight points.

"For the first time, the agent is the deliverable," said Demis Hassabis, chief executive of Google DeepMind. Anthropic and OpenAI are expected to respond within weeks.

— End —