Images 2.6 vs FLUX 3: Edit Fidelity Against Resolution Control

The comparison between Images 2.6 and FLUX 3 is usually framed as a quality contest. It is not one. OpenAI's Images 2.6, shipped on 4 November 2026 as the API pair GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, is selling iteration: fast turns, a subject that survives a chain of edits, and now motion and vector output. Black Forest Labs' FLUX 3 Image, live since 1 October, is selling determinism: ten reference images, bounding-box layout instructions, fixed resolution tiers up to 4K, and a per-image price you can put in a spreadsheet without modelling token counts. One of them is a better product for a design team that needs the same layout every time. The other is a better product for a team that needs to try twelve variants before lunch.

This is the practical version of that choice — what each control surface actually lets you do, what an image costs on each, and what the only audited leaderboard says about the generation before this one.

What each model actually is

Images 2.6 is OpenAI's third image generation in seven months. Images 2.0 arrived in April as the first reasoning image model, Images 2.5 split the line into a speed tier and a quality tier in September, and 2.6 keeps that split: Flare for latency, Sunburst for detail and multi-edit consistency. Both are hosted, both bill by token, and both are reachable through the standard images endpoints. There are no weights.

FLUX 3 Image is Black Forest Labs' flagship, released a month earlier. It handles text-to-image and editing with up to ten reference images, accepts bounding-box layout instructions, and renders at fixed tiers from 768×768 to 4K with selectable aspect ratios. Two weeks after the flagship, BFL published FLUX 3 Image Klein — a 5.6B distilled sibling with downloadable weights and a licence that is free for commercial use under $5M in annual revenue. That open-weights escape hatch has no counterpart anywhere in the Images 2.6 release.

The control surface: where FLUX 3 still wins outright

Four things FLUX 3 does that Images 2.6 does not, or does not document.

• Layout you specify rather than describe. Bounding boxes let you place elements by coordinate. Images 2.6's markup-guided editing is powerful but different in kind: you show the model where to change something on an existing image, you do not hand it a layout grid for a new one.

• Resolution chosen up front. FLUX 3 publishes its tiers — 768, 1K, 1.5K, 2K, 4K — and prices each. Images 2.6's documentation, as published on 5 November, describes quality and complexity settings and output formats but does not publish a resolution table. If your pipeline has a hard 4K deliverable, that asymmetry alone can decide the question.

• Ten references. FLUX 3's multi-reference editing is the most concrete specification in this matchup. OpenAI's release talks about reusable reference images inside ChatGPT and image input on the API, without publishing a per-request ceiling.

• A self-hosted path. Klein's weights are downloadable. Nothing in the Images line is.

Editing: the gap Sunburst exists to close

OpenAI's pitch for Sunburst is not peak beauty. It is that the tenth edit still looks like the first — the face, the logo, the product detail hold. That has been the real failure mode of hosted image models for two years: every edit is a fresh generation conditioned on the last, and small drift compounds.

Two of the 2.6 API features point directly at that problem. The complexity parameter lets you give the model more planning room when a scene has several elements that must all survive; markup-guided edits remove the ambiguity of "change the thing on the left" by drawing on the image instead. FLUX 3's answer to the same problem is structural — feed it more references, and tell it where things go.

Which approach works better is not yet a question anyone can answer with data. Sunburst has no independent edit score. FLUX 3 Image does: 1,123 Elo on Artificial Analysis's editing board, which puts it in the top fifteen but comfortably behind the previous OpenAI generation.

What an image costs on each

The billing models are different enough that a naive per-image comparison misleads in both directions.

• FLUX 3 Image, per image — $0.041 at 768², $0.048 at 1K, $0.07 at 1.5K, $0.10 at 2K, $0.607 at 4K.

• FLUX 3 Image Klein, per image — $0.008 at 768², $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K, $0.036 at 4K.

• Images 2.6, per token — $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), $30 per million image output tokens, identical on Flare and Sunburst.

Worked through, the token rate lands in the same neighbourhood as FLUX 3 at ordinary sizes: roughly half a cent for a low-quality 1K render, about five cents for a high-quality 1K, and around ten cents at 2K high. Where the two diverge sharply is at the extremes. A 4K FLUX 3 image costs sixty cents because you bought a 4K image; a low-effort Images 2.6 render costs almost nothing because you bought very few output tokens. Neither price is a trick — they simply charge for different things.

The number that never appears on either rate card is the cost of a rejected image. If Sunburst's edit consistency saves you one re-roll in five, its effective price per usable asset drops by 20% and the comparison flips. Until there is independent edit data, that is an assumption, not a calculation.

Benchmarks, with the asterisks attached

Artificial Analysis's text-to-image board on 5 November ranks GPT Image 2.5 Sunburst first at 1,224 Elo and GPT Image 2.5 Flare second at 1,208. FLUX 3 Image sits at 1,119, around twelfth, at $48 per thousand images. Klein is at 1,043 for $9 per thousand.

Read that carefully. The generation OpenAI shipped in September beats FLUX 3 by roughly 105 Elo on a board with a ±10 confidence interval — a real gap, not noise. Images 2.6 is not on the board at all yet, so the fair statement is that OpenAI's previous model outscores FLUX 3 on human preference, and the new one is unmeasured. It would be easy to write "2.6 beats FLUX 3"; nobody has run the votes.

On editing, FLUX 3 Image scores 1,123 and ranks in a nine-to-seventeen band. GPT Image 2.5 Sunburst leads that board at 1,199. Again: the predecessor, not the model in the title.

The preference boards also do not measure the things FLUX 3 is actually best at. Nobody voting in an arena is checking whether the bounding box landed on the right pixel or whether the 4K export is genuinely 4K. Treat the Elo as a proxy for "which output looks better to a stranger", which is a real question and not the only one.

Open weights change the risk profile, not the quality

Klein is the most underrated part of this matchup. At 5.6B parameters and nine dollars per thousand 1K images on BFL's own pricing, it is cheap enough to use as a draft tier, and its weights being downloadable means the pipeline survives a price change, a deprecation or an outage. The licence boundary is revenue-based: free commercial use under $5M annual revenue, separate terms above it. Read that clause before you build on it.

On quality, Klein is not a substitute for either flagship — 1,043 Elo is a visible step down from FLUX 3 Image's 1,119 and a long way from the OpenAI numbers. The honest use case is volume pre-visualisation with a flagship finishing pass, not a cost-cut replacement.

Choosing between them

Pick FLUX 3 when the deliverable is specified before generation starts: fixed aspect ratios, a layout that has to repeat, 4K output, a compositing pipeline that wants reference images rather than prose. Pick Images 2.6 when the work is iterative and conversational — the same asset edited twenty times, a character that has to stay on-model, a preview loop where six seconds versus twelve changes what your users do. Pick Klein when you need the model to keep working when someone else's API stops.

If you want to test both without two contracts, OrcaRouter cannot do that today: neither Images 2.6 nor FLUX 3 is in our catalogue yet. What is: GPT-Image-2.5 Flare and Sunburst, Google's Gemini and Imagen image models, and Grok Imagine Image, behind one key with automatic failover and provider list price passed through at 0% markup. That makes a bake-off cheap on the models we do carry, and it means a vendor price change reaches you the same day rather than at the next contract renewal.

The thing to watch is the first independent Images 2.6 score. If Sunburst lands at or above 1,224, FLUX 3's case narrows to control and resolution — which are real advantages, but narrower ones than they look today.