OpenAI Images 2.6 adds GIFs, faster generation and two API models

What OpenAI released

OpenAI announced Images 2.6 on November 4 [1] and made it available in ChatGPT and Codex across subscription tiers, as well as through two API model IDs.

GPT-Image-2.6 Flare is the speed-oriented option. OpenAI says it can generate an image in as little as six seconds when quality is low. GPT-Image-2.6 Sunburst is intended for higher-quality work and for keeping characters, products and other details consistent across edits.[1]

The product can create animated GIFs from prompts or images. The API documentation also describes complexity control, markup-guided edits and SVG output.[2] These are separate controls and formats, rather than proof that every generated asset will meet a production brief.

Flare or Sunburst: a practical choice

Use Flare for thumbnail exploration, mood boards, variant generation and interactive tools where a person is waiting. Use Sunburst for final product scenes, dense layouts, reference-sensitive work and edit chains where unintended changes are expensive.

Keep model, size, quality and complexity explicit in production requests. If you leave complexity unset, the API guide says both 2.6 models default to high. Higher complexity can add latency and can change details outside the requested region, so begin with low for a focused edit and move upward only when the scene needs more planning.[2]

Run the same evaluation set against both models. Include small text, hands, logos, repeated characters, crowded scenes, one-object edits and three consecutive edits to the same source. Score instruction following, reference fidelity, unwanted changes, latency, moderation outcomes and returned token usage.

A migration checklist for image workflows

• Save 20 to 50 representative prompts and the input images used for edits. Include difficult and ordinary cases rather than polished demos only.
• Run Images 2.5 and both 2.6 models with the same size and quality settings. Record response time, token usage and human preference without revealing the model name to reviewers.
• Test complete edit chains. Ask for one precise change at a time and compare all untouched regions after each round.
• Use low complexity for a localized change, then test medium or high only when the model misses relationships or dense layout constraints.[2]
• Validate GIF frame count, duration, loop behavior and file size in every target client. Review each frame because a safe first frame does not guarantee a safe or accurate sequence.
• Treat SVG output as code. Sanitize it, reject scripts and external references, and render it in an isolated preview before publishing.
• Send a stable privacy-preserving safety identifier for end-user requests, as OpenAI recommends, and avoid sending an email address or username directly.[2]
• Keep the previous model available until the new version passes quality, cost, latency and safety thresholds for the actual workload.

Pricing and capacity planning

OpenAI lists both 2.6 models at $5 per million text-input tokens, $8 per million image-input tokens and $30 per million image-output tokens. Cached input has separate rates.[3] These rates match the listed 2.5 rates.

A matching token rate does not make each image cost the same. OpenAI states that models and quality settings can consume different output-token counts.[2] Measure the usage returned by the API and calculate cost from real traffic rather than copying one sample estimate.

The API guide shows 823 output tokens and an estimated $0.02469 output charge for a square low-quality 2.6 image at the time of publication.[2] Text and reference-image input are additional. GIF and SVG workflows can add work or model calls, so measure the complete asset pipeline.

Safety, provenance and production limits

• OpenAI says it moderates prompts, input images and generated outputs. The company also added safeguards for GIF creation, but its system card still reports some unsafe generations in adversarial evaluations.[4]
• OpenAI reports its own speed and quality results. No single benchmark covers a team’s brand, typography, product and edit requirements.[1]
• Generated text, geometry and identity details can still be wrong. Require human review for advertisements, packaging, instructions, public figures and regulated content.
• An editable reference does not grant rights to a person, logo, character or photograph. Check consent, licence terms and local law before publishing.
• SVG files can contain active content. Sanitize generated vectors before they reach a browser, design system or asset repository.
• Moderation can block valid work or allow an output that a team should reject. Keep application-level policy checks and an appeal path.

Frequently asked questions

• Is Images 2.6 one model? The product name covers two API models: Flare for speed and Sunburst for detail and edit consistency.[1]
• Does 50% faster mean every request takes half as long? No. It is OpenAI’s aggregate claim. Prompt, model, quality, complexity, size and service load affect latency.[1][2]
• Did the token prices rise? The published 2.6 rates match 2.5, but a request can use a different number of tokens.[2][3]
• Can the API make GIFs? Yes. The 2.6 image API supports animated GIF output with frame and duration controls.[2]
• Should every 2.5 integration switch immediately? No. Run a matched regression set and compare full per-asset cost, latency and quality before changing the production model.