OpenAI and Anthropic both got worse today — one shipped, one lost a researcher
A Tuesday-night double: OpenAI shipped a new image model before its own math proof has cooled, and an Anthropic pretraining researcher walked out the door saying both labs are gambling with human lives. Different labs, same pressure.
OpenAI has shipped ChatGPT Images 2.5 in two flavours, and the headline number is a 50% cut in generation latency against Images 2.0. The default, GPT-Image-2.5 Flare, is tuned for speed — OpenAI says two to four times faster than its predecessor — and handles transparent backgrounds cleanly, which is the difference between "nice demo" and "actually usable for brand assets and presentation decks." The second model, GPT-Image-2.5 Sunburst, trades that speed for fidelity, aimed at product photography and marketing work where a slightly wrong shade or a softened logo is a rejected deliverable. Both reach ChatGPT, ChatGPT Work and Codex across desktop, mobile and web today, with separate API endpoints for developers.
The genuinely new thing is Sketch, which lets you draw a rough composition directly in the chat window and have the model treat your scribble as a structural guide rather than a prompt. Combine that with inline comments on generated images for targeted edits, templates for flyers and product shots, and the guarantee that earlier edits stay locked in through a long multi-turn session, and OpenAI is describing something more specific than "image generator": a tool for iterative creative work, where the tenth version of an asset still looks like the first one's subject. It ships with C2PA metadata and invisible watermarking, and it lands one day after the same company announced an AI-generated Millennium Prize proof — a reminder of how much surface area one lab now covers.
Jacob Coxon resigned from Anthropic on Monday night, in a post that has since passed five million views. He spent three years in pretraining research at OpenAI and then Anthropic, and his case is unusually specific because he names both employers: at OpenAI, he says, many people have not internalised the stakes; at Anthropic, they understand them perfectly and are racing anyway because they believe no one else will stop. His line — "neither company is acting responsibly... racing straight to self-improving superintelligence and gambling with our lives" — is the strongest version yet of the argument that the race itself, not any individual model, is the risk.
What makes it land differently from the usual departure is that Coxon is not an ethicist or a policy hire; pretraining researchers are the people whose daily work is the capability curve, and they rarely leave saying the curve is the problem. It also reads as a direct answer to Monday's Navier-Stokes announcement: OpenAI told the world it has an internal model "significantly more capable than GPT-6 Astra," and within hours one of the people who builds those models said the pace is the hazard. We covered the capability side of that story in OpenAI's AI solved Navier-Stokes — and a mathematician cries foul.
What to watch: whether OpenAI responds to Coxon's claims at all, and whether Sketch shows up in competing assistants before the year is out.
If the people building these models say the pace is unsafe, what would actually slow it down? Tell us in the comments.
