OpenAI releases Images 2.6 with faster generation and more editing controls

OpenAI released Images 2.6 on November 4, adding another option for teams that generate and revise visual assets. According to the official announcement, it is available globally in ChatGPT and Codex across subscription tiers, with GPT-Image-2.6 Sunburst and GPT-Image-2.6 Flare available through the API.

The practical question is how reliably a useful image survives the next edit. A faster first generation saves little time if each revision breaks the product shape, character details or typography that already worked.

What is available now

OpenAI says generation is 50% faster than before, with Flare reaching as little as six seconds at low quality. Sunburst targets greater detail and consistency at medium or high quality. These are vendor claims, not PromptsZone measurements.

The release adds GIF generation, brings image markup to the API and supports SVG output there. Its new complexity parameter controls planning effort separately from image quality. That distinction gives developers another setting to test rather than one universal “best” configuration.

A safety comparison, not an image-quality leaderboard

No verified, matched multi-benchmark creative-quality comparison against leading competing image models was established for this report. The available comparable numbers below concern safety outcomes. They cannot tell a designer which model produces the best composition, follows typography instructions most accurately or preserves identity across edits.

OpenAI evaluated saved generations from the same 3,077 adversarial requests with common reference grading. Safe-generation rates were 76.15%, 75.50%, 78.36% and 75.04% for 2.5 Sunburst, 2.6 Sunburst, 2.5 Flare and 2.6 Flare, respectively. Unsafe-but-blocked rates were 21.94%, 22.78%, 20.15% and 22.81%. Unsafe-not-blocked rates were 1.92%, 1.72%, 1.49% and 2.14%.

Neither overall unsafe-not-blocked change was statistically significant. Flare’s decline in safe generations and increase in unsafe-but-blocked outputs were significant against its 2.5 baseline after adjustment. The test does not measure the full deployed system or production incident frequency. Bold indicates only the numerical best within each outcome, following the source’s direction of improvement.

These are three outcomes from one evaluation, not three independent benchmarks. A larger share of blocked unsafe generations should also be read alongside how often unsafe content was generated in the first place. Treating any one cell as an overall safety ranking would lose that context.

What teams should test before switching

Editorial evaluation checklist: use the same small set of real tasks on the current model and the candidate replacement. Keep the prompt, input image, output dimensions and quality setting fixed, then record the settings that cannot be matched. Separate quick revisions from complex scenes so an average does not hide failures in either workflow.

• Measure the finished asset. Record time and cost through the final approved revision, including failed attempts, instead of timing only the first image.
• Check what should stay unchanged. Ask for a small local edit, then inspect the rest of the image for unintended differences. Repeat with text, product details and recurring characters.
• Validate exports in their destination. Open SVGs in the actual design or publishing tool. Inspect animated loops frame by frame and at their intended display size.
• Keep human review. Confirm rights to reference material, inspect factual claims and check brand details before anything reaches an audience.

Developers should consult the current API pricing page and record actual usage for their chosen settings. A generation-speed claim alone does not establish a lower cost per approved asset.

Images 2.6 is worth evaluating as a production-workflow update. The release’s tools are concrete; the strongest case for adopting them will come from repeatable tests on the work a team actually needs to finish.

Visual credits: original editorial diagram and data chart, not Images 2.6 output. Both use the official OpenAI provider wordmark as a disclosed fallback, sourced from the displayed wordmark on OpenAI’s brand guidelines. The mark is rendered faithfully with proportional scaling and belongs to OpenAI. Its use follows the linked Marks usage terms and does not imply endorsement. No AI-generated demonstration image or invented benchmark is presented.