🚀 Grok 4.7 Launches Rivaling GPT 5.6 Sol

SpaceXAI dropped Grok 4.7, and this one looks less like a routine model update. The bigger story is how quickly Grok is closing the gap on the frontier while competing hard on price. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: SpaceXAI has launched Grok 4.7, its new flagship for coding, AI agents and knowledge work.


Key Points:

SpaceXAI built Grok 4.7 on a larger base model than 4.6 and trained it longer with reinforcement learning on harder, multi-hour tasks. It is better at checking its work and handling long context, and it now understands the company’s Grok Bot harness natively. The model is available in Cursor, Grok Build and the API with a 500K context window. Though independent tests found higher token usage on hard tasks, lower rates may not mean a cheaper finished job with the new model.

OpenAI and Anthropic now have more pressure on where they can charge a premium. Grok 4.7 is much better at long-running agent work while keeping its API prices low, which gives companies another serious option for coding and knowledge tasks. The real question is what each finished job costs. If Grok can deliver similar work for less overall, more companies may start routing tasks between models based on price and performance instead of defaulting to the most expensive one.

Evolving AI: Artificial Intelligence Live! is built around actually doing AI, not just hearing people talk about it.

From November 15–20 in Orlando, you can get hands-on with AI agents, custom copilots, MCP, vector search, AI security, governance, Azure OpenAI, Semantic Kernel, Copilot Studio, and more.

With 200+ sessions and 30+ deep dives across six co-located conferences, you can build your week around the technologies and skills most relevant to your work, and bring practical knowledge back to your team.

If you've been considering attending, September 25 is the date to keep in mind. Register before the deadline with code “NLSEPT4” and you'll save $400 on your pass.

Save $400 on Artificial Intelligence Live! before September 25 →

Evolving AI: Palantir, Nvidia and Booz Allen all fear the labs could retain or learn from proprietary data.

OpenAI began discussing how to share the model’s growing body of mathematical results with the wider research community. It then approached several mathematicians about creating an external advisory board. Instead, with OpenAI’s agreement, they formed an independent group hosted at Princeton’s Institute for Advanced Study and invited additional members to join. Its current task is to advise on how a large number of significant results from the model should be reviewed and released. Members are unpaid, can publish their recommendations, offer advice without being asked, and also advise other AI companies. Final decisions still remain with OpenAI.

Why It Matters:

OpenAI may be reaching a point where producing important mathematical results is becoming easier than reviewing, verifying and releasing them responsibly. If that continues, the bottleneck around frontier AI could shift from model capability to the human systems around it. Labs may need new structures to decide which machine-generated results matter, how they are checked, who gets credit, and how quickly they enter the research community. The new mathematics group offers one early model for how that oversight could work while keeping researchers outside the company involved. Similar questions could eventually emerge in other scientific fields if AI starts generating results faster than expert communities can absorb them. The biggest caveat is that the full set of 100+ results is still not public, so their real significance has yet to be tested outside the labs.

Evolving AI: British Columbia is suing OpenAI and Sam Altman, alleging the company failed to warn police before a school shooting killed eight people.

OpenAI's safety systems first detected conversations linked to violent activity about eight months before the February attack. The new lawsuit alleges members of its safety team wanted police contacted, but senior leadership decided against it. OpenAI has said its review at the time did not find an imminent and credible threat that met its reporting standard. After the first account was banned, the shooter created another one that OpenAI's repeat-offender system failed to detect until after the shooting.

OpenAI has changed its rules, faced lawsuits from victims, and is now being taken to court by a provincial government. That progression could force AI companies to rethink what happens after their systems detect a serious threat. Refusing a prompt or banning an account may no longer be enough if there is credible evidence of real-world danger. The harder question is when a private conversation crosses the line into something that should reach law enforcement, and who makes that call. This case could push OpenAI, Anthropic and others to build clearer escalation systems around that decision, even as the courts still have to determine whether a police warning here would actually have prevented.

QUICK HITS

💻 Google opened preorders for Googlebook with on-device Gemini and five laptop models starting at $899.

🎨 Qwen released Qwen-Image 2.1 weights with native transparency and 10-image editing, but commercial use needs a separate license.

🤖 Boston Dynamics opened an Atlas training center inside Hyundai’s Georgia plant, where humanoids are learning real factory tasks.

🇨🇳 Alibaba unveiled its new Zhenwu V900 and plans Qwen models with 5 to 10 trillion parameters.

 *partner link