ChatGPT now makes GIFs you can edit frame by frame
Plus: Uber counted four agent-booked rides and 500 food orders during its busiest week in October.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:005:04
Read transcript
ChatGPT can now turn a photo or a text prompt into an animated GIF—and, importantly, let you revise it one frame at a time. OpenAI released Images 2.6 on November 4. The animation works like a flipbook: the model generates related still images, then strings them together. You can inspect individual frames, adjust playback speed, and change a single frame before sharing the result. In a demonstration described by The Decoder, a dancing cat stayed recognizable across the sequence. That suggests the approach can work for that example; it isn’t a measured reliability result across other characters or scenes. OpenAI pitches the feature for things like animated mascots, social reactions, product loops, and storyboards. A sketch can also serve as the starting point. The tools are available globally in ChatGPT and Codex, across subscription tiers and on desktop, mobile, and web. Developers get two API models: GPT-Image-2.6 Flare, positioned for speed, and Sunburst, positioned for more detail and consistency through edits at medium or high quality. OpenAI says the release generates images 50 percent faster than its predecessor. Those are company claims, and neither model appeared on Artificial Analysis’s image-generation or editing leaderboards in OrcaRouter’s November 5 check. The API adds controls for targeted edits and SVG export, too. One useful distinction: a complexity setting controls how much scene-planning effort the model applies; image quality is a separate choice. OpenAI recommends low complexity for most editing, and higher complexity for scenes with many elements or constraints. The broader shift is toward AI visuals that are not just generated, but adjustable—and that puts a premium on human review as these systems take on more complicated work. That emphasis on control has a political counterpart: voters and candidates are also demanding more say over where AI infrastructure goes. Abdul El-Sayed in Michigan and Manny Rutinel in Colorado won congressional seats after campaigning for tighter AI controls. Data centers featured in other races, too. WIRED reports that voters in nine states rejected development plans, approved moratoriums, or sought more community input. Incoming governors in Georgia and Texas backed construction pauses. These results give critics a louder political voice, but they are not a new federal AI policy, and campaign positions are not restrictions already in force. Local ballot decisions, meanwhile, address particular projects or communities, not a single nationwide rule. The practical thing to watch is whether campaign promises turn into enacted pauses, local oversight, or congressional investigations. And in research, the same question of what AI can do—and what still needs checking—shows up in Anthropic’s account of work with Claude. The company says researchers used the model to help resolve the first complete vault structure from prokaryotic life. The shell is made of 96 copies of a protein from a virus that infects bacteria. Its biological purpose remains unknown. Claude helped search about 15 billion predicted proteins for candidates, and supported lab and cryo-electron microscopy work. Anthropic says an optimized preparation method took about 90 minutes from cells to microscopy grid, compared with roughly three days for the conventional method it described. Researchers guided the analysis, and Anthropic says the first full model was ready in 34 hours; the final maps took about a week and hundreds of GPUs in parallel. The process was not error-free: a Claude-written growth protocol included an additive that interfered with DNA uptake, delaying the project by two weeks. The findings were submitted as a preprint, and the shell’s function is still unresolved. The result is a research advance with meaningful human direction and verification—not evidence that the model independently solved the biology. The need to validate AI output is especially concrete in security. OpenAI says a prototype pipeline generated 948 working exploits that let an existing low-privilege Linux user gain root access, plus demonstrations of 291 information leaks. These were local privilege-escalation attacks, not remote break-ins. OpenAI says it tested stock desktop virtual machines, rather than systems deliberately weakened for testing. The system narrowed its review to about 2.8 million lines of kernel code, then generated and tested candidate findings in isolated machines. OpenAI reports that the full run consumed $1,213,430 in API credits. That is not an all-in research-cost figure, and the company says the work prioritized methodology over cost optimization. OpenAI has begun reporting critical flaws, but candidate patches still need review by maintainers. Across image tools, biology, security, and data-center politics, capability is only part of the story. Watch what happens at the handoff: whether people can inspect and revise generated work, verify a research result, approve a security patch, or influence a construction decision. That is where useful AI claims meet real-world accountability.
