GPT-Image-2.5 vs LLaDA-Image: A $30/M-Token API Against a Free 6B Diffusion Model

Ask what image generation should cost and two labs built the two opposite answers within a week of each other. On September 8, 2026 OpenAI shipped GPT-Image-2.5 — the model behind ChatGPT Images 2.5, sold through the API as GPT-Image-2.5 Flare and GPT-Image-2.5 Sunburst — priced on a token card of $8 per million image-input tokens and $30 per million image-output tokens. Five days earlier, on September 4, inclusionAI (the Ant Group-backed lab behind the LLaDA language family) quietly released LLaDA-Image, a six-billion-parameter diffusion-transformer image generator and editor whose weights are downloadable for nothing. One is the most commercially powerful image model OpenAI has ever put behind a token meter; the other is an attempt to prove that a strong image model can be built with an open training recipe and given away. Comparing them is less a head-to-head than a decision about which definition of cost you are actually paying.

Almost every number on the GPT-Image-2.5 side is vendor-reported and none of it has an independent check yet: OpenAI says Flare cuts latency by up to 50% versus GPT-Image-2 and that third-party tests by the agent company Manus measured 2–4× faster generation, but as of September 9, 2026 neither GPT-Image-2.5 model has an entry on the Artificial Analysis image leaderboards. LLaDA-Image's headline figure is equally unreproduced — inclusionAI reports 53.53 English / 53.38 Chinese on Qwen-Image-Bench, first among open-weights models at release — and because the model has no hosted API it cannot appear on API-only leaderboards at all. Both sides of this comparison are running on their makers' own claims, which is worth holding onto through the rest of this piece.

Two models that barely share a category

GPT-Image-2.5 is a hosted, closed image model in the same lineage as ChatGPT Images 2.0 from April 2026. It is multimodal on the input side and produces images that OpenAI claims are sharper in fine detail, more natural in lighting and texture, and steadier across multi-turn edits — the Sunburst endpoint exists specifically for edit-heavy production work where slower generation is acceptable in exchange for tighter instruction control, while Flare is the fast default for high-volume generation. Outputs run up to 2048×2048 and 3840×2160, transparent backgrounds are supported, and every output carries C2PA provenance metadata plus an invisible watermark.

LLaDA-Image is a research-style open release built on a completely different bet. Where GPT-Image-2.5 is served by OpenAI's infrastructure, LLaDA-Image is a set of weights you run yourself: a 6-billion-parameter diffusion transformer (DiT) on the generation side — the model card records about 7B parameters once the language side is counted — paired with LLaDA 2.0-mini, a 16B-total / 1B-active mixture-of-experts diffusion language model, as the "understanding" side that turns text prompts or edit instructions into the signal the DiT follows. The unusual claim in the arXiv report (2609.03796) is the training recipe: more than 90% of roughly 220 million cumulative pre-training and mid-training samples are image-only, with a frozen vision-language model extracting semantics from unlabeled images as a self-conditioning signal, so the model "learns to draw first, then obeys." Progressive resolution takes training from 256×256 through 512×512 up to 1024×1024 fine-tuning, followed by mixed text-to-image and editing training and a TwinFlow few-step distillation that produces the LLaDA-Image-Turbo variant.

What each one is actually good at

The GPT-Image-2.5 pitch is precision and control at commercial quality. OpenAI's announcement stresses reference fidelity — keeping a person, pet, or product recognizable when you change the setting or style — and editing that changes only what you asked to change, sustained across many rounds of edits in one conversation. Text rendering inside images is called out specifically, including the elimination of garbled Chinese characters that plagued earlier image models. Those are exactly the capabilities a product, brand, or advertising team hits repeatedly.

LLaDA-Image's pitch is a single set of weights that does bilingual generation and instruction-guided editing with an open recipe. inclusionAI reports native Chinese and English text rendering, an editing path that injects reference images through a dual-path design meant to preserve unmodified content, and — on Qwen-Image-Bench — the top open-weights scores at release. The honest caveats from the release are that perceptual edit quality still trails specialized editors and counting objects remains weak. It is a generalist open model, not a finished commercial product.

The scoreboard is deliberately not a contest on image quality — the two models have never been shown the same evaluation. What it shows is where the money and the control flow.

The price of one image is the number neither side will give you

OpenAI publishes a token rate card for GPT-Image-2.5 identical to GPT-Image-2's: $8 per million image-input tokens, $30 per million image-output tokens, cached image input at $2, and text tokens billed separately at $5 per million. What it does not publish is how many tokens a given image consumes, which means there is no official per-image price and no cost calculator. At AA's listed price for the predecessor on its leaderboard — roughly $211 per 1,000 images for GPT Image 2 at "high" quality — a token-metered flagship is an expensive tool at volume, but that figure is for GPT-Image-2, not 2.5, and per-image cost for the new model is genuinely unknown outside OpenAI.

LLaDA-Image has the opposite honesty problem. The weights cost nothing and the card carries an Apache-2.0 tag, but the real price is the infrastructure: a 6B diffusion transformer needs a serious GPU for the 50-step Base variant, and the FP8 checkpoints (about 49 GB in a diffusers layout) are the practical option for most users. LLaDA-Image-Turbo's 2–4-step distillation is the concession to that reality. Community tooling has moved fast — an SGLang pull request adds text-to-image, editing, sequence-parallel and FP8 support, and a RebelAI ComfyUI node package supports INT8 and GGUF-quantized local inference — but "free model" still means "you are the hosting provider." For a team that already runs GPU capacity, the marginal cost of LLaDA-Image approaches zero; for everyone else, the all-in cost of serving it can exceed metered API bills once you count engineering time.

The honest path if you want both

These are not either/or for most teams. A production pipeline can use a hosted API like GPT-Image-2.5 for the customer-facing, edit-heavy, must-not-fail work and keep an open model like LLaDA-Image for experiments, offline batch jobs, and anything that cannot send prompts to a third party. That split is exactly the shape of problem a routing layer is built for — one API that reaches the hosted models you actually use, provider list price passed through with no markup, so trying an open-weight model through a hosted route later costs the same as going to the provider directly. Neither model is on OrcaRouter's catalog as of September 9, 2026: GPT-Image-2.5 is OpenAI-API-only for now (the previous generation, GPT-Image-2, is routed), and LLaDA-Image is weights-only with no hosted inference. The design intent stands regardless of today's catalog.

Which one should you build on?

If your work is commercial image generation where a failed edit costs a client relationship — product shots, campaign creative, reference-consistent imagery, long multi-turn edit sessions — GPT-Image-2.5 is the model whose entire design targets that job, and the Sunburst endpoint is the reason to pay attention even before independent scores exist. The risks are the per-image price opacity and the fact that every quality claim is OpenAI's own. If your work is research, high-volume experimentation, Chinese-and-English bilingual generation, or anything that must run in your own environment or on your own data, LLaDA-Image is the more interesting release — genuinely open weights with a published recipe, at the cost of being your own infrastructure and living with unreproduced scores. The models are one week apart in release and a world apart in operating model; the right choice is whichever one matches the cost you are actually able to pay.