GenAI Chronicle
2010
2011
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2025

July

5 events

Editors' summary

Gold-medal standard at the math olympiad, and coding talent moves

On 21 July, Google DeepMind said Gemini Deep Think had solved five of six problems at the International Mathematical Olympiad, reaching gold-medal standard as an official entrant graded by IMO coordinators. OpenAI had announced the same score on the 19th, but was not an official entrant and had its solutions graded by former medallists. Some criticised OpenAI for going first, as the IMO had asked companies to hold off.

On the 11th, Google brought Windsurf CEO Varun Mohan, co-founder Douglas Chen and others into Google DeepMind with a $2.4 billion non-exclusive licence to the startup's technology — a company OpenAI had been in talks to buy for around $3 billion days earlier.

On the 17th, OpenAI released ChatGPT agent, merging Operator's work on websites, deep research's synthesis and ChatGPT itself into one system that picks its own tools as it goes.

On the 22nd it added 4.5 gigawatts to Stargate with Oracle, taking capacity under development past 5 gigawatts.

This block is written by the editors. It is kept separate from the sourced record below.

Record5 events
  1. 11
    BusinessGoogle / Windsurf / Varun Mohan / Douglas Chen

    Google hires Windsurf's leaders in a $2.4 billion licensing deal

    Google brought in Windsurf CEO Varun Mohan, co-founder Douglas Chen and a small group of staff to Google DeepMind, together with a non-exclusive licence to the AI coding startup's technology, in a deal worth $2.4 billion. It was not an acquisition: Google took no stake and Windsurf stayed independent. The agreement followed the collapse of OpenAI's talks to buy Windsurf for around $3 billion, and the arriving team was said to be working on agentic coding for Gemini.

  2. 11
    ModelMoonshot AI

    Moonshot AI publishes Kimi K2, a trillion-parameter open model

    China's Moonshot AI released Kimi K2, a sparse mixture-of-experts model with a trillion parameters in total and about 32 billion active per token, selecting 8 of 384 experts, with a 128,000-token context. The weights shipped under a modified MIT licence. What set it apart was the tilt toward agentic work — using tools and carrying multi-step tasks through on its own — trained by synthesizing large numbers of tool-use scenarios in post-training. The reaction repeated the shock of DeepSeek: near-frontier capability you could download rather than rent. An update in September strengthened coding and widened the context to 256,000 tokens.

  3. 17
    ProductOpenAI

    ChatGPT agent picks its own tools and gets to work

    OpenAI folded an agent into ChatGPT that chooses its own tools as it works. ChatGPT agent merged three things that had been separate — Operator's ability to work websites, deep research's synthesis, and ChatGPT's conversation — into one system. Working on its own virtual machine, it moves between a browser, a terminal and code execution to go from checking a calendar through gathering material to producing editable slides and spreadsheets, pausing to ask the user to log in where a site requires it. It went to Pro subscribers first, with Plus and Team following.

  4. 21
    ResearchGoogle / OpenAI

    Models from two labs reach gold-medal standard at the math olympiad

    On the 2025 International Mathematical Olympiad problems, models from Google DeepMind and OpenAI each solved five of six problems for 35 points out of 42 — the gold-medal threshold. Google DeepMind's Gemini Deep Think was an official entrant, graded by IMO coordinators, and announced on 21 July. OpenAI was not an official entrant; it had its solutions graded by former IMO medallists and announced on 19 July. Some criticised OpenAI for going first, as the IMO had asked companies to hold off. Because the two results were verified by different routes, whether they can be compared side by side remained open to argument.

  5. 22
    BusinessOpenAI / Oracle

    Stargate adds 4.5 gigawatts with Oracle

    OpenAI and Oracle agreed to develop an additional 4.5 gigawatts of Stargate data centre capacity. Together with the Stargate I site in Abilene, Texas, that takes capacity under development past 5 gigawatts, enough to run over two million chips. The companies said construction and operations would create more than 100,000 jobs. It marked progress against the commitment announced at the White House in January to invest $500 billion in 10 gigawatts of AI infrastructure in the US.

↑ 2025