News · 2026-11-05

OpenAI Images 2.6 adds native GIFs and new controls, with token costs worth checking

OpenAI launched Images 2.6 on November 4 with native animated GIF generation, annotation-guided edits and an additional planning control. Its two developer-interface variants carry the same published token rates as the previous generation, but OpenAI’s own documentation shows why that does not guarantee the same cost per finished picture.

Key facts

• OpenAI says Images 2.6 launched across ChatGPT, Codex and the developer interface on November 4.
• Flare targets quicker iteration; Sunburst targets detailed images and consistency across edits.
• Standard image-output pricing is $30 per million tokens, rather than a fixed fee per image.
• The primary source is OpenAI’s launch announcement, which says Flare can respond “in as little as six seconds.”

The release follows Images 2.5’s drawing and editing tools. The most visible addition is a small animation workflow. A user can start from a sketch or character concept, ask for a looping reaction and adjust the result. Developers receive frame-count and frame-duration controls for native GIF generation. The important endpoint limitation is straightforward: generation supports GIF output, while the image-edit endpoint does not.

The release also adds complexity settings alongside image quality. Quality describes visual fidelity. Complexity controls how much planning goes into the scene, with a high default. OpenAI’s image-generation guide says higher complexity can increase latency and sometimes introduce unwanted changes. It is therefore a decision about generation behavior, not a lock that guarantees selected objects remain untouched.

Think of the difference as commissioning a poster. A higher-quality printer can reproduce finer detail. Giving the designer more planning time may instead change where the objects go, which details appear and how the composition works. Both can affect the result, but only one is primarily about the fidelity of the final rendering. Conflating the two settings makes performance comparisons harder to interpret.

The vector-output feature is also a staged process. OpenAI says it creates a reference image and then uses that image to produce static vector content. The quality setting affects the reference. Billing includes both image tokens and the authoring model’s tokens, and a vector file cannot be supplied as an image-edit input. This can be useful output without establishing a general native vector-editing system.

Pricing deserves similar attention to units. The official pricing page lists $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens for the two variants. Those are rates for consumed units. A picture that consumes more units costs more even when the unit price has not moved.

OpenAI’s calculator illustrates the mechanism with a low-quality square example using 823 output tokens, about 2.47 cents for image output alone. That is an example, not a complete fixed invoice: input charges and other workflow costs can apply. The useful comparison is a matched task’s total cost and acceptable output, rather than two models’ advertised rates in isolation. The site’s token-economics lesson explains why that distinction extends beyond images.

Early reactions are mixed. In the Hacker News launch discussion, participants questioned calculator token counts, visual quality and how novel the output formats are. The Reddit testing thread includes both praise and complaints about texture, anatomy, reference adherence and jerky animation. These are individual reports without a controlled comparison or representative sampling.

A same-day hands-on article described higher costs and imperfect loops, but its page could not be opened during dossier verification. Those detailed measurements are not treated as confirmed evidence here. The documentation already supports the narrower economic warning without relying on an inaccessible experiment. OpenAI’s claims of faster generation likewise remain vendor performance claims rather than independently established general results.

There is a security and provenance component. The Images 2.6 system card describes safeguards for harmful meaning that emerges across animation frames. It acknowledges that motion-related violations are difficult to detect. Its challenging offline test set is not a sample of ordinary production requests, so its safety percentages cannot be read as the rate of harmful images users receive.

The system card also adds DeepMind’s invisible SynthID watermark alongside continued provenance metadata. The methods supply different signals, and neither guarantees that every altered copy remains identifiable. That follows the distinction in the site’s provenance lesson. The release’s practical value is new creative output and finer controls; the honest evaluation remains whether a matched workflow produces an acceptable result at an acceptable total cost.

Primary source, verified: read the paper →

Ground Truth. (2026, November 5). OpenAI Images 2.6 adds native GIFs and new controls, with token costs worth checking. Ground Truth. https://groundtruth.day/news/openai-images-2-6-gifs-token-costs.html

@misc{groundtruth:openai-images-2-6-gifs-token-costs,
  title  = {OpenAI Images 2.6 adds native GIFs and new controls, with token costs worth checking},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/openai-images-2-6-gifs-token-costs.html}
}

Topics: image-generation · products · openai · pricing · provenance