August was about people. Google, the company with the most compute and the deepest research bench, reorganized its AI leadership.
DeepMind changes hands
On August 5, Demis Hassabis stepped down as CEO of Google DeepMind to become its chairman. He added the title of chief scientist for Alphabet and will keep running Isomorphic Labs. Koray Kavukcuoglu, DeepMind's CTO, takes over day-to-day operations as a senior vice president reporting directly to Sundar Pichai, and now owns Gemini model development.
The same day, Jeff Dean announced he was leaving after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. It is a public benefit corporation that aims to automate the research process itself. Alphabet is investing.
Fortune reported that the changes followed low morale, departures, and model delays inside DeepMind. Putting Gemini under a leader who reports straight to Pichai reads as a bet on execution speed over research independence. Google has every input it needs: chips, data, distribution, and cash. This year the gap was shipping cadence, and the reorg targets it.
Astra's ten proofs
On August 1, OpenAI said an internal version of Astra, its next major model, had produced ten advances in mathematics and theoretical computer science, including solutions to three more Erdős problems. The proofs are machine-verified in Lean, and OpenAI put the compute cost at about $2,000 at Sol API rates.
The cost is the important number. Problems that stayed open for decades were solved for about $2,000 in compute. When generating research gets that cheap, the bottleneck moves to choosing the right problems and verifying the answers. That is the bet Discovery Loop is making.
Meta enters the coding agent market
Meta launched Muse Code on August 5, a terminal coding agent in beta built on Muse Spark 1.2 and aimed at large codebases. For big tasks it starts several subagents working in parallel in isolated git worktrees, so the main branch stays untouched. Pay-as-you-go pricing starts at $1.25 per million input tokens, with a much cheaper tier for developers who agree to share data with Meta.
Claude Code and Codex have owned this category for a year. Meta is competing on price, and the data-sharing tier makes the trade explicit: cheaper tokens in exchange for training data. Enterprises will avoid that tier. Individual developers may not.
Google also launched Gemini 3.7 Flash and the Pixel 11 lineup this month.
What August tells us
When everyone has money and chips, talent and org design still decide outcomes. And as AI makes research and code cheap to generate, the valuable work shifts to choosing problems and checking answers.