Latest / UpNext AI / ChatGPT Work, Cursor’s OpenAI Cutoff, and AI Essay Scoring | UpNext AI – August 31, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Monday, August 31st, 2026, and here's what matters in AI today. Simon Willison’s practitioner analysis of ChatGPT Work describes a product that has evolved quickly since OpenAI announced it on July 9th. His central finding is that “Work” is really two offerings: a cloud service accessed through ChatGPT, and a desktop version that can access local files and run programs on a user’s computer. The analysis focuses on the cloud product. For paid subscribers starting at 20 dollars per month, Willison found tools that go beyond a standard chat interface: code execution with internet access, a headless Chrome browser that can fill forms and take screenshots, persistent files across sessions, website publishing, and parallel sub-agent sessions. He also found scheduled automations that can be combined with those capabilities. That is a meaningful shift in what a general AI assistant can do. It can move from drafting an answer to carrying out a multi-step task involving code, web pages, and reusable files. But the same combination raises the familiar agent-security challenge: private data, untrusted web content, and tools that can act outward. Willison argues that OpenAI needs to make its protections against prompt injection clearer. For teams considering Work, the useful question is not just whether it can complete a workflow, but which permissions, sites, and data that workflow should be allowed to touch. That expanding role for AI tooling makes a contract dispute involving a major coding assistant worth watching. OpenAI says it has notified SpaceX that it intends to wind down its contract supplying OpenAI models to Cursor, the AI coding tool SpaceX acquired. The company’s proposed shutoff date is November 12th, 2026, and it says it is providing the maximum notice allowed under its contract so developers have more time to retain access through Cursor. OpenAI says the decision was difficult and stresses its respect for Cursor’s team and developer community. Its stated reason is that it cannot be confident SpaceX will use its technology within OpenAI’s terms of service, citing its experience with Elon Musk’s companies and the need for custom agreements with large partners to ensure compliance and safety at scale. The company also points to greater accountability as its upcoming Astra model advances. The immediate practical consequence is a transition problem for developers who depend on OpenAI models inside Cursor. OpenAI has not said what alternatives or capabilities will remain in Cursor after the contract winds down. More broadly, this is a reminder that an acquisition can alter not just ownership of a developer tool, but the model access and commercial relationships beneath it. For the research note, consider a practical question for schools: can AI grade student writing in a way that is reliable and fair enough to support teachers? A new benchmark study in Moroccan secondary education compared four tools against human reference scores on 60 authentic student essays. The researchers used a rubric aligned with the Common European Framework of Reference for Languages, or CEFR. Essays received a global score out of 20, plus six individual criteria. The study tested GPT-4o and Claude 3.7 Sonnet alongside the French-language grammar tools Grammalecte and LanguageTool. GPT-4o had the lowest mean absolute error—an average gap of 3.77 points from the human global score—but it also showed a systematic tendency to underscore essays by 2.68 points. Grammalecte, by contrast, had nearly zero average scoring bias. The models agreed more strongly on formal dimensions such as grammar and spelling than on discursive qualities such as cohesion and register. The takeaway is not that one tool has solved automated grading. The study’s reproducible approach—fixed prompts, a structured output format, and a CEFR rubric—offers schools a way to test systems locally before deploying them. AI scoring may be more useful as a checked input to assessment than as an unexamined substitute for teacher judgment. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... A Cambridge computer scientist and OCaml maintainer reports that public discussion of a security patch can now draw attempted exploits within minutes. In one case, probes for a suspected path-traversal flaw arrived about ten minutes after a patch was shared for discussion. His warning is that open-source disclosure and embargo practices may need to move faster as automated vulnerability discovery improves. Tencent has released Hy4 Preview, an open-weight, text-only model with 770 billion total parameters, of which 49 billion are active at a time. According to Simon Willison’s review, it also has a one-million-token context window. The release illustrates how rapidly the scale of open-weight models is increasing, though the download itself is substantial at 1.56 terabytes. Andrew Bailey, head of the global financial-stability watchdog, has urged G20 countries to take steps to control the release of new frontier AI models. The Financial Times reports that Bailey warned the models could pose dangers to the financial system, putting AI release policy squarely on the agenda for financial regulators. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!