OpenAI has now removed the official DALL·E GPT from ChatGPT. In its ChatGPT release notes, the company said the tool would retire on 30 August 2026 and told users to download any images they still needed. Image generation itself is not going away: OpenAI is pointing people at ChatGPT Images, and user-created GPTs with image generation enabled are not affected.
That is a product-line cleanup more than a capability cut. DALL·E 2 and 3 already left the API earlier this year. The practical implication is for teams that still have old DALL·E GPT threads as the system of record for assets — those galleries will not migrate. Anyone treating ChatGPT as an image archive needs to export today, then rebuild workflows around ChatGPT Images.
Read OpenAI’s ChatGPT release notes.
What else matters
Labs want a common way for agents to drive real hardware
Anthropic opened a research preview of the Model Hardware Standard, a specification for AI agents to operate lab and factory devices such as microscopes, liquid handlers and robotic arms. The work started with HHMI Janelia. Partners can apply while Anthropic writes safety evaluations before an open-source release. If it sticks, the bottleneck moves from model quality to who is allowed to let an agent touch physical kit.
Frontier labs ask for a collective cyber defence
More than a hundred firms — including OpenAI, Anthropic, Google and Microsoft, plus cyber vendors such as CrowdStrike — signed an open letter, as reported by TechCrunch, calling on companies and governments to treat AI-enabled attacks as a shared problem. The letter argues hospitals, water plants and internet infrastructure are at risk as models get more capable. That is advocacy, not a new law: the same labs are still shipping more powerful agents.
Enterprise Q&A gets a harder, larger benchmark
Researchers posted CorporateBench on 27 August: a human-validated Q&A benchmark built around four synthetic firms and corpora of more than 230,000 documents. It tests extraction and knowledge-base querying at a scale closer to real corporate mail and files than typical toy RAG sets. Treat it as a preprint, not a leaderboard fact, until independent labs reproduce the scores.
Caterpillar is treating jobsite AI as a workflow problem
In an interview published today, Caterpillar CTO Jaime Mineart told TechCrunch the company is taking lessons from autonomous mining into construction and quarries. It cites about 1.6 million connected assets and a Cat AI Assistant that lets technicians pull repair procedures by voice. The useful claim is operational, not model-centric: autonomy fails if the site’s process does not change with the machine.
ChatGPT can now bind more than one Google account
On 28 August, OpenAI began letting Plus, Pro, Business and Enterprise users connect multiple Gmail, Calendar and Contacts accounts in one conversation. That is a small product change with a large data-boundary implication: personal and work inboxes can now sit in the same thread, which is convenient and a policy problem for any workplace that still pretends those identities are separate.
