ChatGPT Images 2.6 and the case for local models

OpenAI released Images 2.6 on November 4. The strongest independent results behind the case for using it still belong to 2.5. That is enough to take the new release seriously. It is not enough to call it a tested winner against models running on your own machine.

The local alternatives have a more immediate problem. Two of the stronger downloadable options, FLUX 3 Image Klein and Qwen-Image-2.1, come with conditions that can rule them out before the first image. Choosing between them means deciding what you need permission to do, as well as which output looks better.

I haven’t run a head-to-head test for this article. The comparison below uses published evaluations and the licences for the exact releases, checked on November 6.

The results belong to 2.5

Artificial Analysis ranks images through pairwise human preference votes. On its overall v2.1 text-to-image and editing boards, GPT Image 2.5 Sunburst at max quality was well ahead of these downloadable alternatives:

Evaluated model	Text-to-image Elo	Editing Elo
GPT Image 2.5 Sunburst (max)	1227 ±10	1199 ±11
Qwen-Image-2.1	1045 ±12	1080 ±13
FLUX 3 Image Klein	1043 ±12	1069 ±13
FLUX.2 [klein] 4B	840 ±12	940 ±16

Those are separate scales with reported 95% uncertainty intervals, not percentages. Sunburst’s rows had 13,294 text-to-image samples and 18,575 editing samples. Qwen and FLUX 3 Klein are close enough that their intervals overlap; neither is close to Sunburst here.

This gives OpenAI a stronger quality argument than a gallery of selected launch images. Choosing local control can mean accepting output that people prefer less. But preference votes don’t establish whether an edit preserved a particular product’s geometry or lettering, and neither 2.6 variant appeared on those boards when checked.

OpenAI positions the new API models, gpt-image-2.6-flare and gpt-image-2.6-sunburst, around speed and detail respectively. One useful editing change is separate control over complexity. The guide suggests lower complexity for focused edits, warning that higher settings can introduce unwanted changes. More planning is not always helpful when the request is to change less.

What the download allows

The close results for Qwen and FLUX 3 Klein suggest two plausible local candidates. Their licences make them very different choices for paid work.

Qwen-Image-2.1 documents native RGBA output, subject extraction and editing with up to ten reference images. Transparency and extraction could make it useful for assets that need to sit inside a layout rather than fill the whole frame. Its licence, however, allows research or evaluation only unless you obtain a separate commercial licence. The original Qwen-Image was Apache 2.0. You cannot combine that release’s permissions with 2.1’s results.

FLUX 3 Image Klein also supports up to ten references, with bounding boxes for placement and editing. Its Kommunity licence permits commercial output generation for qualifying users with less than US$5 million in gross annualized revenue, including affiliates. Conditions include content filtering or output review and legally required AI disclosures. That exception is not blanket permission for commercial fine-tuning or redistribution.

For a qualifying small studio generating images, Klein therefore offers a route that Qwen 2.1’s default licence does not. Commercial fine-tuning still needs its own licence check.

The older FLUX.2 [klein] 4B has an Apache 2.0 licence, allowing commercial use subject to its terms. It supports generation and multi-reference editing, with lower preference scores than either newer candidate. Accepting that tradeoff buys broader permissions in a smaller model. Choosing one of the newer models means establishing that its terms cover the work before investing in the setup.

Running it yourself changes the comparison

Even after choosing a licence, downloading the weights doesn’t reproduce a hosted service. Klein’s model card says it expects a full scene caption. Its recommended pipeline uses a separate prompt upsampler to turn a short request into that caption. Sending “make it blue” straight to the image model can reframe the shot instead of making the intended edit.

The same distinction applies to preserving an image. Klein’s bounding boxes are guidance, not hard masks. BFL describes pixel restoration in its hosted API, but that service behaviour cannot simply be credited to the local checkpoint. An API evaluation and a quantized download on a laptop are different systems, even when the model name matches.

Running locally does let you keep the checkpoint and workflow together, including the seed and sampling settings. That makes reruns easier to control, though it doesn’t guarantee identical pixels across hardware or software versions. It can also keep source images off a provider’s servers, provided the upsampler and every other step stay local too. Open weights on a rented GPU are still hosted inference.

Those are concrete reasons to accept a lower preference score. If inputs cannot leave the machine, OpenAI’s lead doesn’t make it eligible. If uploads are allowed and the job is an occasional illustration, the published results give little reason to build a local stack just to avoid using OpenAI.

For recurring work, a local model doesn’t have to beat OpenAI across an entire leaderboard to justify its maintenance. It does need to produce the assets the job requires without giving up the permissions or privacy that made you choose it. That is the test to run once the licence fits.