Making a small animated graphic is becoming one less reason to leave ChatGPT. You can now adjust its speed and edit individual frames right in the chat.
ChatGPT adds moving pictures
ChatGPT’s GIF tools are rolling out alongside Images 2.6. OpenAI’s examples show animation generated from a prompt, with controls to change playback speed and inspect frames; the update also supports starting from an uploaded photo and editing individual frames.
For anyone already making marketing images in ChatGPT, this is a small addition that could save a trip to another tool. A little motion does not need to become a whole production. If the controls have reached your account, try uploading a photo, asking for a GIF, then adjusting the playback speed and checking the frames before sharing it.
Previously, ChatGPT couldn’t natively generate a GIF in one step, Emma Roth writes at The Verge. OpenAI also says Images 2.6 makes image generation faster. The company says people now generate more than 4 billion images a week across ChatGPT and its API, up from the 3 billion images per week it reported in September.
Images 2.6 is rolling out to all ChatGPT and Codex users on web, desktop and mobile.
A smaller model for invoice fields
Dharma Field Extractor, introduced November 4, takes a document image and returns named fields in JSON, a structured format another system can read. Dharma-AI says its 2-billion-parameter model can extract everything it finds or fill an exact set of fields you specify, including invoice line items.
This is for the person connecting document intake to an existing system, likely with technical help. If your recurring chore is pulling supplier names, due dates and totals from invoices, it addresses that particular job. Dharma says missing fields return `null` rather than an invented value.
Treat the savings as a deployment estimate. On its own benchmark of 546 documents, Dharma says it beat all four general-purpose model configurations tested, at running costs roughly 6 to 147 times lower. That comparison uses self-hosted GPU time at $2.24 an hour, not a customer sales price.
Voters put limits on data centers
Local data-center restrictions gained ground in Tuesday’s elections. Lauren Feiner writes at The Verge that, as of Wednesday morning, Redbud Consulting’s tally showed restrictive results in 33 of 45 local ballot measures.
For readers buying AI services, this is background for evaluating vendors’ expansion promises, rather than a reason to change tools today. The results concern where infrastructure can be built and what conditions communities impose.
The politics weren’t uniform: some winning candidates opposed moratoriums, and Feiner cautions against attributing candidate victories directly to data-center backlash. One concrete result came in El Dorado, Kansas, where voters approved a 10-year ban on data centers with peak electrical loads above 10 megawatts.
Also
- Siri Recaps in beta. Sarah Perez found that Apple’s conversation summaries captured repair details but mixed them with household chatter.
- Offline speech with Phonon. Gradium says its English-only Phonon 31m model runs on-device and is free for commercial use under Apache 2.0.
- Copilot inside Xcode. GitHub’s built-in assistant integration is in public preview starting with Xcode 27.3 beta and supports all Copilot plans.
- Agents barely booked Uber. Uber says AI agents booked just four rides during Halloween week, alongside 500 Uber Eats orders.
Pass it on