July 30th 2025

1) Frontier model releases & core research breakthroughs

• Anthropic launches Claude 4 (Opus 4 & Sonnet 4) — May 22. New flagship reasoning/coding models with agentic upgrades (parallel tool use, memory, extended thinking) ship across Anthropic’s products and partner clouds.
• Anthropic activates ASL‑3 (AI Safety Level 3) protections for Opus 4 — May 22. A step‑up in security/deployment standards (especially around CBRN misuse and model‑weight security) applied as a precaution with the Claude 4 launch.
• Google DeepMind shows an advanced Gemini (Deep Think) reaching gold‑medal performance at the International Mathematical Olympiad — July 21. An end‑to‑end natural‑language approach solved five of six IMO problems within the 4.5‑hour limit, indicating a notable leap in formal reasoning.
• xAI unveils Grok 4 — July 9–10. xAI’s new top‑end model rolls out to API and “SuperGrok” tiers; Oracle also advertises Grok 4 on OCI.
• Mistral releases Voxtral (open‑weights audio models) — July 15. Voxtral Small (≈24B) and Mini (≈3B) handle direct speech+text input, long audio context, and multilingual ASR/understanding; SDK support for audio/transcription landed the same day.
• Meta’s super‑sized Llama 4 “Behemoth” delayed — May 15. Reporting indicated a pushback of Meta’s largest next‑gen model, affecting the open‑source frontier‑model roadmap.

2) Policy, governance & safety milestones

• EU publishes the final General‑Purpose AI (GPAI) Code of Practice — July 10. A voluntary but detailed path to comply with the AI Act’s transparency, copyright, and safety/security obligations for GPAI models (with systemic‑risk guidance). The AI Act’s GPAI rules start August 2, 2025.
• Google says it will sign the EU GPAI Code — July 30. Alphabet’s Kent Walker announced Google will sign while flagging concerns about innovation and trade‑secret exposure; major peers’ positions diverge.
• United States releases the “America’s AI Action Plan” — July 27. The White House outlines national priorities spanning safety, competitiveness, energy, workforce, and research coordination.
• Anthropic signals continued work with governments and safety bodies (mid‑July). The company noted intent to sign the EU Code and announced a DoD collaboration on responsible AI in defense operations.

3) Compute & infrastructure (the “AI factories”)

• OpenAI + Oracle add 4.5 GW to Stargate build‑out — July 22. The expansion pushes total announced capacity above 5 GW across sites, signaling an unprecedented AI‑compute ramp to support future models/agents.
• NVIDIA supply dynamics & next‑gen platforms (May–July). Blackwell‑family systems (B200/GB200) begin ramping during 2025H2 with GB300/“Blackwell Ultra” sampling; U.S. regulators allow certain Nvidia H20 shipments to China in late July.

4) Major product changes that reshape how people use AI

• OpenAI’s ChatGPT Agent — announced July 17. A more autonomous agentic mode for ChatGPT, previewed ahead of DevDay 2025.
• OpenAI DevDay 2025 announced — July 23 (for Oct 6). OpenAI tees up its third developer conference, promising previews of “what’s coming next.”
• OpenAI introduces Study mode in ChatGPT — July 29. A pedagogy‑informed mode that guides learners step‑by‑step (Socratic prompts, scaffolding, knowledge checks), rolling out to Free/Plus/Pro/Team users (Edu “in the next few weeks”).
• Google Search AI Mode expands (June–July). Google’s June AI roundup added broader Gemini 2.5 availability, NotebookLM public sharing, and more; on July 28 AI Mode began rolling out in the UK as a tabbed, multimodal “most powerful” search experience.
• Microsoft ships Copilot Mode in the Edge sidebar — July 28. A new “no‑tabs” browsing and task mode where Copilot handles reading, clicking, and summarizing within the sidebar; Microsoft also pushed new Copilot for Microsoft 365 features in late July.
• Apple’s WWDC25 brought new “Apple Intelligence” capabilities — June 9–13. Systemwide features (e.g., image understanding, on‑device live translation, a new Workout Buddy coach) and an expanded on‑device foundation model for developers; watchOS 26 public beta emphasizes subtle AI assistance on‑device.

5) Robotics & embodied AI

• Amazon‑backed Skild AI debuts “Skild Brain,” a general‑purpose robotics model — July 29. Targeting multi‑robot, multi‑environment skills (navigation, manipulation, robustness) with continual data network effects across deployments.
• Google highlights progress toward on‑device robotics models (June recap). Gemini Robotics On‑Device and related research point to more dexterous, generalizable, low‑latency robot control.

6) Corporate strategy moves

• OpenAI merges Jony Ive’s io Products team into OpenAI — update posted July 9 (initial letter May 21). Ive’s LoveFrom assumes deep design/creative responsibilities across OpenAI, indicating a push toward new AI‑native devices/experiences.
• Microsoft–OpenAI extend long‑term access arrangements — June 4. The companies updated their agreement governing model access and infrastructure collaboration.

Honorable mentions & context

• NotebookLM upgrades (video overviews, public sharing) and Gemini 2.5 family expansion continued Google’s push for agentic/workflow tools (early–mid July recaps).
• xAI’s Grok platform rolled out additional features (e.g., “companions”) alongside Grok 4; the release followed moderation incidents that drew scrutiny.

What this means (quick take)

• Reasoning & agents: Claude 4, Grok 4, and Gemini Deep Think pushed long‑horizon reasoning and tool‑using agents from demos toward mainstream products, with education‑ and workflow‑centric UX (e.g., Study mode, AI Mode) starting to adapt around them.
• Compute realism: The Stargate 4.5 GW expansion and Blackwell/GB‑series ramp underscore that infrastructure scale remains a decisive bottleneck/advantage.
• Regulatory clarity: The EU’s GPAI Code lands just before Aug 2 obligations begin for GPAI providers, with early signatories signaling a shift from debate to implementation.
• Embodied AI: Skild’s “shared brain” approach and Google’s on‑device robotics work suggest 2026 could be an inflection year for practical, general‑purpose robots in industrial and service settings.

‍