OpenAI announced Images 2.6 on November 4, 2026, with availability in ChatGPT, Codex and the API. The release adds image-generation and editing controls, including API image markup, SVG export, a complexity setting and image-quality controls. For QA teams using Codex to create UI concepts, test-data imagery or visual assets, this makes one rule more important: generated pixels are an input to validate, not an approved baseline by default.
What OpenAI announced
OpenAI says Images 2.6 is available globally in ChatGPT and Codex across subscription tiers on desktop, mobile and web. In the API, GPT-Image-2.6 Flare and Sunburst offer choices intended for faster generation or higher-detail work. The announcement also describes markup-based edits, SVG export, and controls for visual quality and planning complexity.
Those are product capabilities, not a guarantee that two generations will produce identical assets. If an AI-produced image enters a test environment, QA should preserve the exact asset and the evidence used to approve it.
OpenAI Images 2.6 visual regression QA: a practical plan
- Separate creation from acceptance. Generate or edit assets in a staging workflow. Promote a reviewed file, rather than a prompt, into the visual-regression baseline.
- Record the rendering contract. Store the source asset hash, dimensions, colour profile, prompt/edit instruction, generation date and reviewer decision with the test run.
- Use stable fixtures in CI. Download and version the approved image in the repository or controlled test-data store. Do not regenerate it during a screenshot comparison.
- Test responsive and failure states. Run the asset through the breakpoints, dark/light surfaces, crop rules, lazy-loading path and image-error fallback that customers actually encounter.
- Make diffs reviewable. Keep the reference, actual image and a thresholded diff. A pass/fail number alone rarely explains whether a change is a desirable edit, compression drift or a UI defect.
A small acceptance matrix
| Check | Evidence to retain |
|---|---|
| Approved asset is rendered | File hash and deployment identifier |
| Layout survives each breakpoint | Reference and actual screenshots |
| Edits did not change protected areas | Region-specific visual diff |
| Fallback path is usable | Network-blocked screenshot and accessibility check |
Why this matters for QA engineers
AI image tools can speed up prototypes and test-content preparation, but they add variation to a pipeline that visual testing normally expects to be deterministic. Treating the approved output as a versioned fixture keeps the CI signal meaningful. Treating every new generation as a fresh candidate prevents an attractive visual change from silently becoming a production regression.
Common mistake: comparing changing inputs
A fragile workflow regenerates the hero image, then compares the resulting page screenshot to yesterday’s baseline. The test cannot distinguish model variation from a CSS regression. Instead, approve a specific output first; then use that immutable file for the UI test. Regenerate only in a deliberate content-review change, with a new expected screenshot.
Bottom line
OpenAI Images 2.6 expands the image workflow available in Codex. The safest QA response is modest and repeatable: capture provenance, freeze approved assets, test rendering behaviour independently, and require a human to review material visual diffs. That approach makes OpenAI Images 2.6 visual regression QA useful without confusing generated content with test evidence.