OpenAI Images 2.6 targets the revisions that make AI pictures useful

Changing the date on an event poster should be a small job. If the revision also changes a face, a logo or the rest of the text, someone has to check and repair the whole design again.

That is a useful way to approach OpenAI’s Images 2.6 release: ask how well it helps someone finish a visual, including the revisions, rather than judging only the first picture.

OpenAI announced Images 2.6 on November 4 and says it is available globally in ChatGPT and Codex across subscription tiers, as well as through two API models. It reports faster generation and better preservation of characters, products and brand details during editing. Those are provider claims, not Bright test results. OpenAI’s announcement

What is available, and what has been demonstrated

The API offers GPT-Image-2.6 Flare for fast iteration and GPT-Image-2.6 Sunburst for detailed visual work. Their exact model identifiers are gpt-image-2.6-flare and gpt-image-2.6-sunburst. Both generate and edit images from text and image inputs. Flare documentation · Sunburst documentation

OpenAI’s launch examples include animated characters and brand imagery. These are selected demonstrations. They do not establish how often an ordinary user will get a suitable result, or how much checking and correction it will require.

The distinction matters for Bright’s assessment: the company has announced a deployed product, while the evidence for its practical benefits here remains developer documentation and demonstrations. Bright has not independently confirmed access across accounts or run a verified Images 2.6 test.

The controls worth understanding

Make a specific change. Markup-guided API editing lets a developer supply annotations that identify the intended revision. This capability also supports certain GPT Image 2.5 models; it is not exclusive to 2.6. Markup was already available in ChatGPT. API release notes · Launch announcement

Separate planning from finish. The API’s complexity setting accepts low, medium or high and defaults to high. It controls planning, separately from visual quality. Lower complexity can suit focused edits. Higher complexity may add latency and unwanted changes, so it deserves comparison rather than automatic selection. Image-generation guide

Choose the handoff format. The Image API documents SVG and animated GIF generation. SVG supplies static vector content, with no scripts, external resources or embedded raster images. GIF requests allow 2–64 frames, subject to size and pixel limits. Both formats use the generation endpoint; the image-edit endpoint does not output them. Format documentation

Those API details should not be read as a promise that every equivalent control appears in ChatGPT or Codex. The release announcement establishes product availability; the guide specifies the developer interface.

Three practical uses to evaluate

These are proposed uses, not outcomes Bright has measured.

A poster that needs one correction. An organizer could try changing an event date while keeping the illustration, spacing and other words fixed. The acceptance test is simple: does the corrected date appear, and did anything else change? Compare the entire result with the original, including text that was supposed to stay untouched.

A consistent set of illustrations. A teacher or small publisher could explore several scenes featuring one fictional character. Keep a reference sheet with the character’s colors, clothing and distinguishing features. Review every scene against it. Visual similarity alone does not establish that a teaching diagram is factually correct.

A graphic that survives the next tool. Someone building a website could try a simple SVG illustration, then open it in the intended editor and resize it. Someone making a short loop could inspect the GIF frame by frame and at normal playback speed. Check the exported file, not just its preview. A usable handoff should remain legible, preserve the intended shapes and behave as expected wherever it will be shown.

For a first attempt, use an invented subject rather than a confidential client design or an identifiable person. That keeps the exercise focused and avoids sharing material unnecessarily.

Try a small revision before a whole campaign

Here is an original Bright exercise readers can adapt. It is a test plan, not a report of a completed test.

1. Create a simple square poster with an original illustrated blue bird. Specify the exact title “COMMUNITY GARDEN DAY,” the date “12 MAY,” and a cream background. Ask for no other text.

2. Save the first result and note the model or product used. Check the wording before continuing.

3. Ask: “Change only the date from ‘12 MAY’ to ‘19 MAY’. Preserve the bird, title, colors, layout and background.” Use markup if the interface provides it.

4. Compare the before and after at full size. Record whether the date is correct, every unintended change, and any further corrections needed.

5. Count total time to an acceptable final file, including inspection and repair. If comparing models or settings, keep the prompt and criteria fixed and retain unsuccessful attempts too.

One successful revision would establish only that the particular attempt worked. Several attempts across different kinds of artwork would provide a more useful picture of reliability. For a real event, check the final date, accessibility and other details before publication.

Read the speed and cost claims carefully

OpenAI says Images 2.6 generates images “50% faster than before,” but the announcement does not name the comparator or provide enough test conditions to reproduce that figure. Its six-second Flare example is a best-case claim at low quality, not a typical wait time or an end-to-end design-time saving. Launch announcement

The API model pages list the same standard token rates for both variants: $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens, with separate cached-input rates. That is not a fixed price per image. Flare pricing · Sunburst pricing

SVG also incurs charges for the reference image and the model that authors the vector file. The guide warns that text, consistency and exact placement can still be difficult. Costs and limitations

For a small team, the useful measure is the cost of a finished, checked asset: successful generations, discarded attempts and human correction together. Faster previews can help. The stronger result would be a person getting the visual they need with less total effort, while retaining control over what it says and shows.

Source note: AI-assisted reporting based on OpenAI’s November 4 announcement, API release notes and product documentation checked November 5, 2026. All performance claims remain attributed to the provider. No independent benchmark reproduction, safety audit or firsthand Images 2.6 trial is claimed.