A week ago, the story of artificial intelligence seemed settled. Anthropic and OpenAI were pulling ahead with ever more capable models, while Google looked stuck, Meta looked lost, and even the impressive systems emerging from China appeared to be chasing a lead they could not close. Then, in a matter of days, the script was torn up. Google's reversal was the more surprising because the signs of decline had seemed so clear.

Gemini 3, released late last year, was formidable at launch but was soon overtaken, and later updates were merely incremental. Jeff Dean left to found a startup, taking several researchers with him. Demis Hassabis, the Nobel laureate who built DeepMind into a scientific powerhouse, stepped back from day-to-day leadership, and some teams were moved from London into Google's California core. Many observers read this as a company choosing commercial products over frontier research.

On Tuesday, Google answered with Gemini 4 Argon, which it says matches or beats the flagship models of OpenAI and Anthropic on numerous benchmarks, although the value of benchmarks is increasingly disputed. The pricing may matter more than the scores: during an introductory period, Argon costs up to 80 percent less per token than its rivals, and it stays significantly cheaper afterwards. Meta's arrival was arguably more dramatic. Despite spending billions, it had never established itself at the frontier, and now it claims a model that can compete with the best.