AI Roundup July 2025: Grok 4 Goes Big, Kimi K2 Drops Open Weights, ChatGPT Gets an Agent

TL;DR

July 2025 was the month agents stopped being a demo and open weights stopped being a sideshow. xAI launched Grok 4 with autonomous tool use, Moonshot dropped Kimi K2, a trillion-parameter open-weight model that runs circles around most closed coders, and OpenAI gave ChatGPT its own virtual computer to actually go do things. Meanwhile the EU AI Act's general-purpose AI obligations crossed the finish line, so the regulators showed up too.

xAI Ships Grok 4 and Calls It "The World's Most Powerful Model"

On July 9, xAI launched Grok 4 in a late-night livestream, with Elon Musk billing it as the most powerful model on the planet. The headline benchmarks backed up at least some of the swagger: a new state of the art on ARC-AGI v2, and the Grok 4 Heavy variant became the first model to clear 50 percent on Humanity's Last Exam. Heavy does it by spinning up multiple reasoning agents in parallel at inference time and letting them compare notes.

The part builders should actually care about is native tool use. Grok 4 was trained to drive a code interpreter and a web browser on its own, picking its own search queries and pulling in real-time web and X data. It ships with a 256K context window and is available through the xAI API for anyone who wants to wire it into their own stack.

The catch, as always with xAI, is that the model is inseparable from the platform it lives on. Build on Grok and you build on top of X's data pipes and X's moderation drama, which is a feature for some use cases and a liability for others.

Moonshot Open-Sources Kimi K2, a Trillion-Parameter Agentic Model

On July 11, Beijing-based Moonshot AI released Kimi K2, an open-weight mixture-of-experts model with one trillion total parameters that activates just 32 billion per token. It comes in two flavors: a base model and an instruction-tuned Kimi-K2-Instruct, both under a modified MIT license. This is one of the largest open-weight models anyone has shipped, and it is explicitly tuned for code generation and agentic workflows, meaning it picks its own tools, checks results, and adjusts without you re-prompting it.

Two technical details stand out. First, the sparse activation: 32B active out of a trillion means you get frontier-ish quality without paying frontier compute on every token, which is exactly the tradeoff self-hosters want. Second, Moonshot pre-trained it on 15.5 trillion tokens using a new optimizer recipe called MuonClip with, in their words, zero loss spikes. Stable trillion-parameter training is genuinely hard, and they made it look routine.

The honest caveat: "open weights" does not mean "runs on your gaming rig." A trillion-parameter model, even a sparse one, is a serious infrastructure commitment. But for anyone wanting a top-tier agentic coder they fully control, K2 is the drop of the month.

OpenAI Gives ChatGPT Its Own Computer

On July 17, OpenAI launched ChatGPT agent, which fuses the old Operator browser-control project, deep research, and the core model into one thing that operates its own virtual computer. It comes with a visual browser, a text browser, a terminal, direct API access, and connectors into apps like Gmail and GitHub. You can hand it a fuzzy multi-step task ("analyze three competitors and build me a slide deck") and it shifts between reasoning and clicking until it is done.

Why it matters for builders

• It is metered, not magic. Pro users get 400 agent messages a month, Plus and Team get 40. That pricing tells you the per-task compute is real, and it sets expectations for what agent economics look like in production.
• Computer use is now table stakes. Between this and Grok 4's native tools, the bet across every lab is the same: the model that can act beats the model that can only answer.
• The security surface just expanded. An agent with terminal access and your Gmail connected is a different threat model than a chatbot. Treat it like a junior employee with credentials, not an autocomplete.

The EU AI Act's GPAI Rules Cross the Finish Line

On July 10, the European Commission published the final General-Purpose AI Code of Practice, the voluntary playbook for complying with the AI Act's obligations on general-purpose models. It is split into three chapters: Transparency and Copyright apply to every GPAI provider, while the Safety and Security chapter only bites for models with systemic risk, defined as those trained above 10^25 FLOP, a club of maybe 5 to 15 companies worldwide.

The timing is the point. The GPAI obligations themselves became applicable on August 2, so July was the last full month before the rules went live. Signing the code is the path of least resistance to demonstrate compliance and reduce paperwork. If you are shipping a model into the EU, or building on one that is, this is the regulatory baseline you now operate against.

Google Pushes Veo 3 to Pro Subscribers in 150-Plus Countries

Google spent July widening access rather than dropping a single bombshell. The most builder-relevant move: it expanded its state-of-the-art video model Veo 3 to Google AI Pro subscribers across more than 150 countries and shipped a photo-to-video feature so you can animate a still image into a clip. DeepMind also quietly put out research-grade models, including AlphaEarth Foundations, a kind of virtual satellite that fuses Earth-observation data into a single representation of the planet's land and coastal waters.

None of this is as flashy as a new flagship LLM, but the Veo 3 rollout matters: generative video is moving from "locked beta you beg for" to "included in a consumer subscription in most of the world." That is the phase where it shows up in real production pipelines.

Key Takeaways

• Agents went from demo to default. Grok 4 and ChatGPT agent both bet the house on models that take actions, not just emit text. If you are not designing for tool use, you are building for last year.
• Open weights hit the trillion-parameter tier. Kimi K2 proves the open ecosystem can ship frontier-class agentic models, even if you need real hardware to run them.
• Sparse MoE is the self-hoster's friend. 32B active out of a trillion is the architecture pattern that makes giant open models even thinkable to deploy.
• Regulation is now a build constraint, not a someday problem. With the EU's GPAI obligations live as of August 2, transparency and copyright posture are part of your model's spec sheet.
• Generative video quietly went mainstream. Veo 3 landing in a consumer subscription across 150-plus countries is the signal that AI video is ready for actual workflows.