GitHub made Moonshot AI's open-weight Kimi K2.7 Code generally available in Copilot, calling it the first open-weight model offered as a selectable option in the Copilot model picker.

The model is hosted by GitHub on Microsoft Azure — requests never route through Moonshot — and is billed at provider list pricing under usage-based billing. It began rolling out to Copilot Pro, Pro+ and Max plans across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, the Copilot cloud agent and github.com, with Business and Enterprise to follow. For those enterprise tiers GitHub left it off by default, requiring an administrator to enable the Kimi K2.7 policy, and advised admins to review open-weight models against their own security, compliance and data-governance requirements first.

The UN Independent International Scientific Panel on AI released its Preliminary Report, the first global, independent scientific assessment of AI, produced by 40 experts drawn from all five UN regions and co-chaired by Yoshua Bengio and Maria Ressa.

Its central warning is that current safeguards cannot keep pace with the growth of AI capabilities, and that policymakers face an evidence dilemma: by the time the science is clear, it may be too late to act. The report covers AI trajectories, societal applications in science, health, education and agriculture, economic implications, security and environmental impacts, human rights and democracy, and child safety, and singles out agentic systems as a governance step change. Findings were presented at the inaugural UN Global Dialogue on AI Governance in Geneva, ahead of a full report planned for 2027.

Anthropic released Claude Sonnet 5, making it the default model for Free and Pro plans and available on Max, Team and Enterprise, in Claude Code and on the API as claude-sonnet-5.

Anthropic positions it as closing most of the gap with Opus 4.8 on agentic work at a lower price: introductory API pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026, then standard $3/$15. The model is built to plan, drive browsers and terminals, and run autonomously on multi-step tasks, with adjustable effort levels and a refreshed tokenizer. TechCrunch reported Anthropic citing 63.2% on an agentic coding benchmark versus Opus 4.8's 69.2% and Sonnet 4.6's 58.1%.

Google-backed Apptronik unveiled a nearly 90,000-square-foot robot training facility in Austin called Robot Park, developed with Google DeepMind, where fleets of humanoid robots run logistics, manufacturing and retail tasks to generate real-world training data.

The company also introduced Apollo 2, its latest humanoid, which has operated for more than a year as its data-collection platform and ships in both bipedal and wheeled configurations. Robot Park supplies data to Gemini Robotics under Apptronik's research partnership with Google DeepMind. CEO Jeff Cardenas described the setup as a factory that produces robots alongside a factory that produces data, said Apptronik has built hundreds of Apollo 2 units, and said real production versions arrive in 2027 and beyond.

Meta's FAIR team published Brain2Qwerty v2, a non-invasive pipeline that decodes typed sentences from magnetoencephalography recordings with no implant.

Trained on roughly 22,000 sentences from nine volunteers who each spent about 10 hours typing inside an MEG scanner, it reaches 61% average word accuracy and 78% for the best participant, versus roughly 8% for prior non-invasive methods — Meta reports a 29% character error rate for MEG against 65% for EEG. The architecture combines a convolutional encoder over raw brain signals, a transformer, and a character-level language model, with fine-tuned LLMs supplying semantic context and AI agents used to iteratively refine the decoding pipeline. Meta released the full training code for v1 and v2 and its partner BCBL released the v1 dataset.