Edited from four posts published in July 2025, covering June 30–July 27.
July was the month the race for AI talent and the race for proof ran side by side. It opened with Mark Zuckerberg folding Meta’s AI work into Meta Superintelligence Labs under Alexandr Wang, and it closed with two laboratories claiming gold-medal performance at the International Mathematical Olympiad. In between, Meta hired Apple’s head of foundation models and set out plans for computing clusters measured in gigawatts; Google agreed a reported $2.4 billion arrangement to bring Windsurf’s leaders into DeepMind; Thinking Machines raised $2 billion before shipping a product; and Perplexity was valued at $18 billion. At Microsoft, Satya Nadella addressed job cuts that had passed 15,000 that year, while the company spent about $80 billion on AI infrastructure.
Governments set out rival agendas. Brussels published its voluntary Code of Practice for general-purpose models. Washington issued America’s AI Action Plan, with an order against “woke AI” in federal procurement. In Shanghai, Premier Li Qiang proposed a world AI cooperation organization. And a cloned voice impersonating the US secretary of state showed what the same technology could do in the wrong hands.
The month’s releases tested the distance between a claim and its verification. Microsoft reported that an orchestrated set of models diagnosed 85.5 percent of difficult published cases; xAI released Grok 4 with built-in tool use; OpenAI combined its browsing and research tools into ChatGPT agent. Moonshot released the trillion-parameter Kimi K2 as open weights on the same day OpenAI postponed its own open-weight model. The IMO results drew the sharpest line: OpenAI’s was graded by former medalists it engaged, Google DeepMind’s by the competition’s own coordinators.
Competition results kept coming: in September an advanced Gemini reached gold-medal level at the ICPC programming finals, and DeepSeek’s R1 research passed peer review at Nature. The open-weight contest that Kimi K2 sharpened brought OpenAI’s gpt-oss in August and Kimi K2 Thinking in November. The Action Plan’s science agenda reappeared in November’s Genesis Mission, and the impersonation of Marco Rubio was cited when members of Congress proposed tougher penalties for AI-assisted fraud that month. Meta’s Hyperion cluster became part of a $600 billion US investment pledge in November, and the EU code was followed by the start of obligations for general-purpose models on August 2.
Mark Zuckerberg reorganized Meta’s AI work under Meta Superintelligence Labs. His staff memo named Alexandr Wang chief AI officer and placed Nat Friedman alongside him, leading AI products and applied research.1
The reorganization brought research and product development under new leadership, with superintelligence named as the goal.
Microsoft reported that its MAI-DxO orchestrator, paired with OpenAI’s o3, correctly diagnosed 85.5 percent of cases in a benchmark built from 304 complex New England Journal of Medicine case records. A comparison group of physicians, working without colleagues, reference materials or AI assistance, averaged 20 percent on the cases they completed.2
The results came from a research benchmark of published cases; testing in clinical care was a later step.
Cloudflare introduced pay per crawl in private beta. Website owners could allow a crawler free access, block it, or set a price per request. The system used HTTP’s Payment Required response to communicate pricing and supported payment through Cloudflare.3
The experiment gave publishers a way to manage automated access to their pages and to charge for it.
Sakana AI described AB-MCTS, a search method implemented in its TreeQuest library. It can generate new candidate answers, refine promising ones and choose among models. Initial experiments combined o4-mini, Gemini 2.5 Pro and DeepSeek-R1-0528 on ARC-AGI-2 tasks.4
The reported result measured whether a correct answer appeared among many attempts within a substantial computation budget, showing what collaboration among models could add at inference time.
Safe Superintelligence announced that Ilya Sutskever had formally become chief executive and Daniel Levy president. Daniel Gross had left the company on June 29. Sutskever said the technical team would continue to report to him.5
The announcement settled the leadership of a laboratory that had released no products and described its work only in general terms.
Colin Murdoch, president of Isomorphic Labs, the Alphabet company spun out of DeepMind in 2021, told Fortune that “the next big milestone is actually going out to clinical trials” and that the company was staffing up for them. Isomorphic was designing its own drug candidates in oncology and immunology and had announced $600 million in its first outside funding round on March 31.6
Clinical trials would test the company’s candidates in people.
Bloomberg reported that Ruoming Pang, who led Apple’s foundation-models team of about 100 people, was leaving for Meta Superintelligence Labs on a package reportedly worth millions of dollars a year.7
The move was among the most prominent in Meta’s recruiting of senior AI researchers from its rivals that summer.
An internal State Department cable, reported by the Washington Post, said that someone using AI software to mimic Marco Rubio’s voice and writing style had contacted a US governor, a member of Congress and three foreign ministers. The impostor used a Signal account opened in mid-June and left voice and text messages.8
The case showed how voice cloning could be used to impersonate officials who rely on messaging apps.
xAI released Grok 4 with native tool use and live search. The model could use a code interpreter and web browsing during reasoning. The company also introduced Grok 4 Heavy through a new subscription tier.9
Tool access broadened what the system could attempt beyond answering from stored knowledge; its usefulness depended on choosing suitable tools and checking what they returned.
The European Commission published the final General-Purpose AI Code of Practice. This voluntary tool addressed transparency, copyright, and safety and security, with the last chapter aimed at providers of models posing systemic risk.10
Signing the code offered one route toward compliance with the AI Act’s binding rules for general-purpose models, which were due to apply from August 2.
Google and Windsurf announced that chief executive Varun Mohan, co-founder Douglas Chen and selected researchers would join Google DeepMind. Reuters reported a $2.4 billion arrangement that included a non-exclusive technology license, without Google taking a stake in Windsurf.11
The arrangement combined recruitment with access to technology while Windsurf continued as a separate business.
Moonshot AI released Kimi K2 in base and instruction-tuned variants. Its mixture-of-experts architecture contained one trillion parameters, with 32 billion activated for each token. The release included downloadable weights and support for tool use.12,13
Access to the weights let developers deploy and adapt the model under its license, and put a Chinese model near the front of the open-weight field.
Sam Altman said OpenAI would delay its planned open-weight model, due the following week, to run additional safety tests and review high-risk areas, adding that “once weights are out, they can’t be pulled back.” He gave no new date; it was the second delay.14
The same day Moonshot released its weights, the decision showed the two companies weighing openness and risk differently.
Mark Zuckerberg outlined plans for large AI computing clusters. Prometheus was expected to come online in 2026; Hyperion was designed to expand toward five gigawatts over several years. Reuters reported his statement that Meta would invest hundreds of billions of dollars.15
The plans tied model development to electricity, construction and long-term capital commitments, measured in capacity still to be built.
Thinking Machines Lab, founded by former OpenAI executive Mira Murati, announced a $2 billion financing round led by Andreessen Horowitz. Participants included Nvidia, Accel, ServiceNow, Cisco, AMD and Jane Street, according to the company’s legal adviser.16
The round showed how much investors would commit to a new laboratory before it had released a product.
At the Pennsylvania Energy and Innovation Summit, Senator Dave McCormick’s office announced more than $90 billion in investment commitments spanning data centers, energy infrastructure and training. Its fact sheet listed projects from companies including Blackstone, CoreWeave and Brookfield.17
The commitments, from many companies over different time horizons, linked AI’s growth to power generation in one state.
Google made Gemini 2.5 Pro and Deep Search, which runs many searches and compiles a cited report, available in AI Mode to US subscribers of its AI Pro and Ultra plans. It also began rolling out a feature that lets Search call local businesses to check prices and availability.18
Search was taking on research and errands on the user’s behalf, with the most capable model reserved for paying subscribers.
OpenAI introduced ChatGPT agent, combining capabilities from Operator and deep research. The system used a virtual computer with browsers, a terminal and other tools to carry out multi-step tasks. Users could interrupt its work, and sensitive actions required confirmation.19
The release brought information gathering and action into one workflow, with the user kept in the loop for consequential steps.
Bloomberg reported that Perplexity had raised $100 million at an $18 billion valuation, extending a round that had valued the company at $14 billion a few months earlier. The report cited a person familiar with the matter; Perplexity declined to comment.20
The valuation reflected investors’ bet on AI-driven search as an alternative to Google’s.
OpenAI researcher Alexander Wei reported that an experimental model solved five of six International Mathematical Olympiad problems, scoring 35 out of 42. He said three former IMO medalists independently graded each proof and reached agreement on the scores.21,22
The model was unreleased and the grading was arranged by OpenAI rather than the competition; the team published its proofs for others to inspect.
Google DeepMind announced that an advanced version of Gemini Deep Think solved five of six IMO problems and scored 35 out of 42. IMO coordinators graded and certified the solutions; the announcement included confirmation from IMO president Gregor Dolinar.23
The model wrote its proofs in natural language within the competition’s time limits, and official grading made the score directly comparable with those of human contestants.
The White House released America’s AI Action Plan, setting out more than 90 proposed federal policy actions under three pillars: innovation, infrastructure, and international diplomacy and security. Priorities included data-center permitting and exports of US AI technology.24 President Trump signed three executive orders the same day, among them “Preventing Woke AI in the Federal Government,” which bars federal agencies from buying AI models that compromise accuracy for ideological goals.25
The plan joined technology policy to industrial capacity and foreign policy; most of its measures awaited action by agencies.
At I/O Connect India in Bengaluru, Google announced onshore processing for Gemini 2.5 Flash, new agentic tools in Firebase Studio, and a Google Play–Unity training program. It also highlighted Indian startups building models with Gemma.26
The announcements addressed the practical conditions for adoption: where data is processed, how applications are built and who has the skills to develop them.
In a memo to employees, Microsoft chief executive Satya Nadella addressed job cuts of more than 15,000 positions since January, about 9,000 of them in July, while the company was spending about $80 billion on AI infrastructure in its fiscal year. He called it “the enigma of success in an industry that has no franchise value.”27
The memo set out openly the trade-off between investment in AI infrastructure and the size of the workforce.
Researchers introduced ASI-Arch, a system that proposed neural-network architectures, implemented them and evaluated them through experiments. The original paper reported 1,773 experiments focused on linear-attention architectures.28
The work extended automation into the design of models themselves, within one family of architectures.
The World Artificial Intelligence Conference opened in Shanghai for three days. At the opening ceremony Premier Li Qiang proposed a world AI cooperation organization, observing that global AI governance remained fragmented.29 ASEAN secretary-general Kao Kim Hourn also spoke.30 Geoffrey Hinton warned that AI agents might seek more control to reach their goals, comparing humanity’s position to keeping a tiger cub as a pet.31
Three days after Washington’s plan, the conference put China’s case for multilateral governance alongside technical demonstrations.
Sources checked October 26, 2026. Dates distinguish announcements, reports and later updates.