Introducing Images 2.6: OpenAI Puts Motion Behind the API

OpenAI shipped Images 2.6 on 4 November 2026, and the demo it chose was a waving mascot. That is the version of this launch most coverage will carry: ChatGPT can now make a looping GIF from a prompt or a photo, adjust its playback speed, step through the frames and fix one. The version developers should care about is quieter. The same release put two new API models on the wire — GPT-Image-2.6 Flare, tuned for latency, and GPT-Image-2.6 Sunburst, tuned for detail and for holding a subject steady across a chain of edits — and added four controls that change how you write an image call rather than just how fast it returns. Images 2.6 is live in ChatGPT, in Codex and through the API on every ChatGPT tier, on web, desktop and mobile.

This piece is about what actually changed, what it costs, and which of the launch claims have any evidence behind them yet. The short answer to the last question: almost none, because the model is a day old.

What shipped on 4 November

Five things, in rough order of how much they alter a production workflow.

• Animated GIF output. A still becomes a short loop. In ChatGPT you can set playback speed, inspect individual frames and edit them before sharing. On the API it is an output format alongside PNG, JPEG and WebP, available on both new models.

• A complexity parameter. Low, medium or high, and it is not a quality dial. It governs how much planning the model does before it draws. OpenAI's developer guidance is to start low for ordinary edits and raise it only when a scene carries several elements or constraints that have to survive together.

• Markup-guided editing. You send the source image plus a marked-up overlay showing exactly where the change belongs. The overlay is a transparent PNG matching the first image's dimensions, and it cannot be combined with a mask in the same request. This is the ChatGPT "draw on the picture" gesture, promoted to a parameter.

• SVG export. Vector output straight from the image call, which matters for logos, icon sets and anything headed for a print or design pipeline that would otherwise re-trace a raster.

• Speed. OpenAI claims roughly 50% faster generation overall. Flare at low quality is quoted at about six seconds per image — the number that makes a GIF preview feel interactive rather than batch.

ChatGPT also gained reusable reference images, so a recurring character, pet or product does not have to be re-uploaded on every turn. None of this is a new foundation model in the architectural sense. It is a capability release on the Images line that started with Images 2.0 in April 2026 and split into two tiers with Images 2.5 in September.

The two API models are a routing decision, not a marketing split

GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst are priced identically. They are not the same model with a different default. Flare is built to return quickly and, per OpenAI, also improves layout quality for posters and product imagery — the compositional cases where text and objects have to sit somewhere specific. Sunburst is built for detailed compositions and for the thing that has quietly become the hard part of image generation: keeping a subject recognisably itself through ten successive edits.

That split is the single most useful fact in the release, because it maps onto two different jobs that used to share one model. A thumbnail generator, a storyboard sketcher or an in-chat preview wants Flare. A brand asset pipeline that edits the same hero image across fifty SKUs wants Sunburst, and will happily pay the extra seconds.

Both are text-and-image in, image out, on /v1/images/generations and /v1/images/edits. Neither is a general chat model, and neither accepts video.

What it costs, in the only units that matter

OpenAI bills these models per token, not per picture, and the rate card is unchanged from Images 2.5:

• Text input — $5.00 per million tokens, $1.25 cached.

• Image input — $8.00 per million tokens, $2.00 cached.

• Image output — $30.00 per million tokens.

Because output tokens scale with the number of pixels the model actually paints, the per-image cost is a range, not a number. At 1K, a low-quality render lands near half a cent and a high-quality one near five cents; push to 2K and high quality reaches roughly ten cents. The same token rate that looks expensive next to a flat per-image competitor looks ordinary once you notice you are paying for resolution you asked for rather than an average across a tier.

That structure is also why an unchanged rate card can still be a price cut in practice. If Images 2.6 reaches the same quality in 50% less wall-clock time, and the token count per image is comparable, the cost of a given output is flat while the cost of the pipeline around it falls. Batch workloads that queue images overnight will not notice. Anything interactive will.

OrcaRouter does not carry Images 2.6 at the time of writing, so the honest availability line is OpenAI's own API and several third-party platforms. What our catalogue does hold today is the previous generation, GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst, alongside Google's Gemini and Imagen image models and xAI's Grok Imagine Image — all on one key, at provider list price with 0% markup, with automatic failover between providers. When a model lands here, the vendor's price is our price on the same day; we do not add a routing margin on top of it.

What we cannot tell you yet

There is no independent score for Images 2.6. We checked Artificial Analysis's text-to-image and image-editing boards on 5 November and neither lists GPT-Image-2.6 Flare or Sunburst; the top of the text-to-image board is still the previous generation, GPT Image 2.5 Sunburst at 1,224 Elo and GPT Image 2.5 Flare at 1,208. That is not a knock on the new model. It is simply what "one day old" looks like on a preference arena that needs thousands of pairwise votes before it will publish a number.

The one ranking OpenAI can point to is Design Arena, where it says Sunburst took first place and displaced its own predecessor. Treat that as a vendor-cited arena result rather than an audited benchmark: the claim is specific and checkable, but Design Arena has not published an Elo for the new model that we could verify, and arena position is a preference measure, not a fidelity measure.

The claim with the most substance behind it is the least glamorous one. OpenAI says the service now generates more than 4 billion images a week across ChatGPT and the API, up from 3 billion in September. That is a usage figure, not a quality figure, and it is self-reported — but a 33% jump in seven weeks is the kind of number that explains why the speed work got prioritised over another jump in peak fidelity.

The naming collision worth knowing about

Microsoft shipped MAI-Image-2.6 in August 2026. It is a different model from a different company, it has its own API surface through Microsoft's platform, and it already carries independent scores — 1,168 Elo on Artificial Analysis's text-to-image board and 1,151 on editing. If you search for "Images 2.6 benchmarks" this week, a meaningful share of the results will be about that model, or will blend the two.

The distinction that matters when you are reading a spec sheet: OpenAI's line is Images as a product name and GPT-Image-2.6 as the API identifier. Microsoft's is MAI-Image-2.6. They do not share weights, pricing or a leaderboard row. Anyone quoting an Elo for "Images 2.6" today is almost certainly quoting one of them for the other.

Who should move, and who should wait

If your pipeline is already on GPT-Image-2.5 and the bottleneck is interactive latency, Flare is a straightforward swap and the one-day-old risk is low: same endpoints, same token rates, same output shapes, a documented speed gain. If your bottleneck is edit drift — the fifth edit that quietly changes a face or a logo — Sunburst is the interesting one, and it is worth a bounded test rather than a migration, because the consistency claim is exactly the claim that has no independent evidence yet.

If what you actually wanted was the GIF feature, note the caveat early reviewers hit: the rollout was gradual, and some accounts did not have it on day one. GIFs are also not accepted as an input format on the API — you can produce one, not feed one back in.

The thing to watch over the next fortnight is the Artificial Analysis row appearing. When GPT-Image-2.6 Sunburst gets an Elo, the useful comparison is not against the open-weight field but against GPT Image 2.5 Sunburst's 1,224 — because that is the number OpenAI has to beat to justify calling this a new generation rather than a tuning pass.