OpenAI released Images 2.6 on November 4, 2026, adding animated GIF generation to ChatGPT, Codex and the API alongside two new image models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst. The headline feature is motion: users can turn a prompt, a sketch or an uploaded photo into a short looping animation, adjust its speed, edit individual frames and share it. For anyone building companion characters or adult-adjacent creative work, the more consequential detail is buried in the accompanying system card — OpenAI built a separate safety layer specifically for motion.
What actually shipped
Per OpenAI's launch post and the release materials, the update bundles several things:
- GIF generation in ChatGPT and the API, from text prompts, sketches or existing images, with frame-level editing and speed controls.
- Two API models — GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst — with adjustable complexity settings.
- SVG output in the API, plus markup-guided editing for supported GPT Image 2.5 and 2.6 models.
- Reference-image workflows and faster generation, which OpenAI and third-party coverage put at up to 50% faster image creation.
OpenAI's own framing leans on mascots, characters and product shots — "animating a brand mascot or character," in the company's words. That is exactly the use case companion and character-art users have been hacking together with video models for the past two years.
The part companion developers should read twice
The Images 2.6 system card describes a two-stage filter: LLM-based policy checks run before a request reaches the image model, and generated outputs are checked again before display. That structure is not new. What is new is a GIF-specific evaluation set built around risks that only appear across frames.
OpenAI states that its GIF safeguards address "sexualized movement" and harmful meanings that may not be apparent in a single still image, and that GIF enforcement may be broader than still-image enforcement because motion can imply disallowed content over time. In plain terms: a frame that passes on its own can still fail as a sequence. A static character portrait that clears the filter may be rejected once it loops.
The system card reports 77.7% overall recall on a targeted adversarial test set for GIF safety, while acknowledging that motion-related violations remain difficult to detect and that safeguards will not catch everything. That number matters. Roughly one in five adversarial attempts in OpenAI's own test set got through — and adversarial test sets are, by design, harder than ordinary user behavior, but they are also the closest proxy anyone outside OpenAI has for how the filter behaves under pressure.
What this means for adult content
Nothing in the Images 2.6 materials suggests OpenAI has moved on adult sexual content. The usage policies still prohibit sexualization of anyone under 18 and CSAM outright, and the API image-generation guide still describes all prompts and outputs as filtered under the content policy, with a moderation parameter offering auto as the default and low as a less restrictive setting.
So the practical read for companion and NSFW-adjacent builders is unchanged at the top level and tighter at the edges. OpenAI has not opened a door. It has added a second lock on a door that was already closed — and that lock now watches time, not just pixels.
Why the motion filter is the real story
Every major image platform has spent three years tuning classifiers on single frames. Animated output breaks that model. A looping GIF is a sequence of individually innocuous frames whose meaning lives in the transition between them, which is precisely the gap OpenAI is now trying to close with a dedicated evaluation set and broader enforcement.
Expect the same pattern to spread. Once one frontier lab publishes a motion-specific safety evaluation with a recall figure attached, competitors shipping animation features — and every companion app that generates looping character art — inherit the benchmark as the de facto standard. Developers building on the API should assume GIF requests will be rejected more aggressively than still-image requests for the same prompt, and should plan fallback paths accordingly.
Images 2.6 is available now in ChatGPT, Codex and the API.
Key facts at a glance
- Images 2.6 launched November 4, 2026, across ChatGPT, Codex and the API.
- Two new models: GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst.
- GIF generation supports speed adjustment and per-frame editing.
- OpenAI reports 77.7% recall on a targeted adversarial GIF safety test set.
- GIF enforcement may be broader than still-image enforcement due to motion-based risk.