OpenAI released GPT Image 2.6 on November 4 with two distinct API models — Flare and Sunburst — alongside GIF output, SVG export, and a new complexity parameter. It also started a countdown: three older image model IDs retire on December 1, 2026. If you’re calling gpt-image-1-mini, gpt-image-1.5, or chatgpt-image-latest in production today, you have three weeks.
The December 1 Deadline Is Real
OpenAI is retiring three model IDs on December 1, 2026:
gpt-image-1-minigpt-image-1.5chatgpt-image-latest
The recommended replacements are gpt-image-2.5-flare or gpt-image-2.5-sunburst — 2.6 variants work too. There’s no migration fee; it’s a model ID swap. But that doesn’t mean it’s invisible: output differences are real, and OpenAI explicitly recommends testing before switching. The checklist is straightforward: inventory your codebase for the deprecated IDs, run your production prompts against the new models, compare outputs and costs, then migrate. Leave enough runway — December 1 is a hard cutoff.
Flare vs. Sunburst: Not a Price Tier
The first question every developer asks is: “which one is cheaper?” Neither. Flare and Sunburst charge identical rates — per million image input tokens, 0 per million image output tokens — the same as the 2.5 generation. The split is about workflow stage, not budget.
| Flare | Sunburst | |
|---|---|---|
| Speed | 2.2–3.2× faster | Slower |
| Low-quality generation | ~6 seconds | Longer |
| Edit consistency | Good | Excellent |
| Best for | Drafts, rapid iteration | Final output, branded sets |
The practical workflow: use Flare to establish direction quickly, then switch to Sunburst when you’re refining the final asset. Running consistency-critical edits — recurring characters, product revisions, branded visual sets — on Flare will produce visible drift across rounds. That’s not a bug; it’s what Flare trades for speed. See the full API guide for Flare and Sunburst for deeper integration examples.
GIF Output: What Works and What Doesn’t
GPT Image 2.6 adds animated GIF generation via the output_format: "gif" parameter on the generations endpoint. The specs: 2–64 frames (default: 9), 20–2,000ms per frame (default: 200ms), max 720×720 for 16-frame GIFs.
result = client.images.generate(
model="gpt-image-2.6-sunburst",
prompt="A round robot mascot waves hello, looping animation",
size="512x512",
quality="medium",
output_format="gif",
frame_count=12,
frame_duration_ms=100,
)Three things to know before relying on this in production. First, GIF generation doesn’t stream — you wait for the entire file. Second, you can’t upload an existing GIF for editing; the model only animates content it generates. Third, the system card reports 77.7% moderation recall on adversarial prompts, with motion-related violations the hardest to catch. If GIFs touch anything sensitive in your use case, add a human review step.
Use cases that work well: brand mascot loops, meme reactions, sprite sheets, and sequential storyboards. Keep prompts explicit about what stays fixed frame-to-frame; the model doesn’t infer continuity automatically.
New Parameters Worth Knowing
Complexity is new and separate from quality. It controls how much planning the model does before rendering — not visual fidelity. Options are low, medium, high (default: high), with no extra charge. Use low for single-element edits; use medium or high for dense scenes with multiple constraints.
Image Markup lets you direct edits with a transparent PNG annotation layer. Draw a circle around an element, describe the change, and the model targets that region. It works on the edits endpoint with both 2.5 and 2.6 models.
SVG export is available on the generations endpoint only — not edits — and only with n=1. Watch for the billing quirk: SVG output charges twice, once for the reference image model and once for the authoring model rendering vector code. Identical token rates don’t mean identical cost per request when SVG is involved.
What to Do Now
If you’re not using any of the three retiring model IDs, there’s no immediate pressure — but GPT Image 2.6 is effectively a free upgrade on existing token pricing. The speed improvement on Flare (2.2–3.2× faster) reduces latency without increasing cost. The complexity parameter gives you a free knob to dial back over-planning on simple edits.
If you are on the deprecated IDs: start the migration now. Test on staging with your actual production prompts, not toy examples. The model outputs are not guaranteed to be identical, and December 1 leaves no room for surprises. Check the OpenAI pricing page to baseline your cost expectations before and after the swap.
