OpenAI announces Images 2.6: preserve details, target edits and read GIF and cost conditions

Imagine a shop using a cat with a yellow scarf in its promotional material. This time, it wants a room background replaced with a beach rather than a newly designed character. Supplying the original, the features to preserve and the area to change makes the intended difference easier to describe.

Image markup points to an editing area with marks on the image. OpenAI says this existing ChatGPT capability is coming to the API. Saved references help bring material into a later chat; the face, scarf and lettering still need comparison between the original and edited file.

%% Comparing an imaginary shop's edit · request diagram, not generated output
flowchart TD
  A["Before: cat in a room<br/>Yellow scarf is a reference detail"] --> B["Person: mark the background<br/>Request the face and scarf stay"]
  B --> C["Images 2.6: produce an edit<br/>Change the background to a beach"]
  C --> D["Person: compare with the original<br/>Review scarf, face and wording"]

The API has two separate controls. complexity adjusts how much the model plans to meet the request; quality controls rendered detail and fidelity. Changing one background and arranging many objects and words require different kinds of planning. Official image generation guide

The guide suggests considering low complexity for focused edits and comparing higher levels for dense scenes. Omitting complexity defaults to high on both models. A higher setting can increase latency or introduce unwanted changes, so it should not be treated as a single bundled switch that also raises or lowers quality.

The announcement's as-little-as-six-seconds claim is for Flare at low quality. The guide says complex prompts may take up to two minutes; six seconds is not a completion promise for every image.

Imagine making a GIF of the shop cat waving. You could ask in a ChatGPT conversation, or have the shop’s app send a request to OpenAI through the API.

A GIF animation shows several pictures at chosen intervals and repeats them. For a waving cat, inspect both the changes between frames and the transition from the end back to the start. This is an imagined movement design, not generated frames.

The Image API accepts 2 to 16 frames per GIF. A frame is one picture in the sequence. In a hypothetical sequence of nine frames shown for 200ms, or 0.2 seconds each, one play lasts 9 × 0.2 seconds = 1.8 seconds. This calculates playback length, not generation time.

%% An imaginary wave joins back to its start · picture-sequence explanation
flowchart TD
  A["Start: cat's paw is down"] --> B["Middle: raise the paw to wave"]
  B --> C["End: return to the starting pose"]
  C --> D["Replay from the first picture<br/>Review the join for a sudden jump"]

The API guide lists image output at $30 per million billable tokens for Images 2.6. Tokens measure billable usage; a token is not an image. The next calculation assumes a hypothetical output of 1,000 tokens. Official pricing

SVG represents a picture with shapes and lines rather than pixels. The guide says SVG output incurs both the image model's reference-image cost and the authoring model's token cost. Complexity and markup have no separate surcharge, but actual usage still needs a record.