OpenAI's $3 billion agreement to acquire AI coding startup Windsurf fell apart after its exclusivity period expired, and Google DeepMind moved within hours to hire Windsurf CEO Varun Mohan, co-founder Douglas Chen, and several senior researchers to advance its agentic coding work. Bloomberg reported Google is paying about $2.4 billion covering a non-exclusive license to certain Windsurf technology plus compensation for the hires — no equity stake and no control of the company, which remains free to license the same technology to Google's competitors. Windsurf's head of business, Jeff Wang, took over as interim CEO, with most of its roughly 250-person team staying put.

The structure is the now-familiar reverse acquihire — the same pattern Google used to bring back Character.AI's Noam Shazeer and Microsoft used for Mustafa Suleyman — which lets acquirers absorb talent and a technology license without triggering the merger review a full deal would. The stated cause of the OpenAI collapse is the Microsoft relationship: Microsoft holds rights to OpenAI's IP and could not absorb Windsurf's coding technology too. Windsurf's ARR had reportedly reached about $100 million by April, up from roughly $40 million months earlier.

METR published a randomized controlled trial in which 16 experienced open-source developers completed 246 real tasks on large mature repositories they had roughly five years' prior experience on. Each task was randomly assigned to allow or forbid AI usage; developers forecast AI would cut their time by 24%, but the measured effect was a 19% increase in completion time. The same developers, afterward, still estimated AI had made them about 20% faster — a roughly 39-point gap between belief and measurement.

The researchers stress the result is setting-specific and do not conclude that AI fails to help most developers, noting the tools were early-2025 vintage and mostly Cursor Pro with Claude variants. Their candidate explanations are technical rather than psychological: more time spent prompting and waiting, and AI faring worst on large, complex codebases — exactly the environments senior engineers work in.

The European Commission received the final version of the General-Purpose AI Code of Practice, a voluntary framework drafted by 13 independent experts with input from more than 1,000 stakeholders, designed to let model providers demonstrate compliance with the AI Act's GPAI obligations. It has three chapters: Transparency and Copyright, which apply to all GPAI providers, and Safety and Security, which applies only to providers of models with systemic risk. The AI Act's GPAI rules enter into application on 2 August 2025, becoming enforceable by the AI Office one year later for new models and two years later for existing ones.

Signing is voluntary but offered as a lighter-touch compliance route: providers who adhere gain reduced administrative burden and greater legal certainty versus those who prove compliance by other means. Providers deploying GPAI models into the EU — including open-weight ones, whose provider status is the live question — need to read the Copyright chapter's policy requirements and the Model Documentation Form closely.

xAI announced Grok 4 via a livestream, claiming state of the art among closed models and introducing Grok 4 Heavy, a parallel test-time-compute variant that runs multiple agents on a problem and reconciles their answers. xAI says Grok 4 Heavy is the first model to score above 50% on Humanity's Last Exam, at 50.7% on the text-only subset, with native tool use, real-time X and web search, a 256,000-token context window on the API, and a code interpreter trained via reinforcement learning at pretraining scale on the 200,000-GPU Colossus cluster. Access split into a consumer tier and a new $300-a-month SuperGrok Heavy plan, with API access available immediately.

The launch landed one day after the earlier Grok model posted a string of antisemitic responses on X, including praise for Hitler — an incident xAI attributed to prompts and said it had reverted. That sequencing matters more than the benchmark sheet: the same model lineage being marketed on reasoning capability and simultaneously failing basic content boundaries in public is the failure mode enterprises cite when they refuse to expose agents to end users.