Images 2.6 vs HY Image 3.5: Four Cents Against Ten, and What That Buys

Run the arithmetic before the argument. A hundred thousand 2K images on Tencent's HY Image 3.5 costs about $2,400 at its published rate of $0.024 per image. The same run on OpenAI's Images 2.6 — the API pair GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, released on 4 November 2026 — lands somewhere near $10,000 if you model a high-quality 2K render at roughly ten cents, and considerably more if your prompts are heavy and your outputs are large. That is a fourfold gap at the conservative end and closer to ninefold if you compare against the measured $210.70 per thousand that Artificial Analysis records for OpenAI's previous quality tier at its maximum setting.

So the question is not which model is better. It is whether the difference between them is worth four to nine times the invoice — and the answer depends on a distinction the launch coverage has mostly missed: these two models are good at different things, and only one of them is finished.

What Tencent actually shipped

HY Image 3.5 preview went live on 22 September 2026. It is a unified text-to-image and image-to-image model built on the 80B mixture-of-experts Hy Image 3.0 base, with multi-turn conversational editing as a first-class mode rather than an afterthought. It accepts up to five reference images per request, renders up to 2K, and is billed only on output — $0.024 per 2K image on Tencent Cloud International, ¥0.15 on the domestic cloud.

The distribution is the tell. Tencent put it into Yuanbao, ima, Miora, WorkRally, OnSolo and WorkBuddy, and into TokenHub, MPS and VOD on the cloud side, with third-party access through GMI Cloud and ComfyUI. That is a model built to be consumed inside a consumer and enterprise product suite, at a price that makes per-user image generation a rounding error. It is not a premium API product with a design-tool story.

The word "preview" is in the name for a reason. There is no published SLA, no deprecation policy, and a full version is expected to follow. Anyone putting it in a production path is accepting that the endpoint may change under them.

Where HY Image 3.5 genuinely wins

Three things, and they are not small.

• Price at volume. At $0.024 per 2K image, a million images costs $24,000. The equivalent OpenAI run is a six-figure line item. For catalogue rendering, avatar generation, social asset factories and anything where the image is one of thousands rather than one of ten, this dominates every other consideration.

• Text inside images, especially non-Latin text. Tencent's model is explicitly tuned for multilingual rendering and small-text accuracy, and independent reviewers have found it competitive on dense labels and packaging copy. OpenAI's Images line is strong on short English headlines; the further you get from that, the less certain anyone should be.

• Conversational multi-turn as the native mode. HY Image 3.5 was designed for a chat surface where a user says "make it warmer, move the logo left, now try it in blue". Images 2.6 has caught up on the feature list — reference images and frame editing in ChatGPT — but Tencent's product has been shaped around that loop since launch.

Tencent's own blind-test figures put it about 30% ahead of Hy Image 3.0 and roughly level with Seedream 5.0 Pro, slightly ahead of Nano-Banana Pro and Qwen-Image-3.0 Pro. Those are vendor-reported GSB results, not third-party scores, and should be read as a directional claim.

Where Images 2.6 wins

• Headroom. HY Image 3.5 stops at 2K. OpenAI's model does not publish a resolution table, but the Images line has consistently gone higher, and the 2.6 release adds SVG output — vector, not another raster size. If your deliverable is print, 4K, or a design file that will be rescaled, the comparison ends here.

• Speed at the low end. Flare at low quality is quoted at about six seconds per image, with generation overall roughly 50% faster than the previous generation. Tencent publishes no comparable latency figure, and a preview endpoint is not where you go for throughput guarantees.

• Motion and vector formats. Animated GIF output and SVG export exist on 2.6 and have no equivalent in HY Image 3.5. If the deliverable moves or scales, this is not a quality comparison at all.

• The editing ceiling. HY Image 3.5 scores 1,092 Elo on Artificial Analysis's editing board — respectable, and better than its text-to-image number. But OpenAI's previous generation, GPT Image 2.5 Sunburst, leads that same board at 1,199. The gap is real and it is the gap the new model is trying to widen.

The scores, and the missing one

Artificial Analysis lists HunyuanImage 3.5 (Preview) at 971 Elo on text-to-image, around sixtieth, and 1,092 on editing, in the high teens to mid-twenties band. Its measured cost is $24 per thousand images. That is a low score next to the frontier — GPT Image 2.5 Sunburst leads text-to-image at 1,224 — but it is a real score with a sample base behind it.

Images 2.6 has none. Both boards were checked on 5 November 2026, the day after launch, and neither GPT-Image-2.6 Flare nor Sunburst appears. The only ranking OpenAI cites is Design Arena, where it says Sunburst took first place ahead of its own predecessor. That is a vendor-cited arena result; we could not verify an Elo for it.

This is the awkward symmetry of the matchup. The cheap model has an audited score and no resolution headroom; the expensive model has resolution headroom, new formats and no audited score. Anyone who tells you which is better today is extrapolating from the previous generation of one of them.

The hundred-thousand-image test, done properly

Take a realistic catalogue job: 100,000 product images, 2K, one reference image each, short English prompt, no edits.

• HY Image 3.5 — 100,000 × $0.024 = $2,400. Fixed, predictable, no token modelling.

• Images 2.6 — output tokens scale with painted pixels, so the same job is roughly $10,000 at a high-quality 2K equivalent, and materially less if the images are simple enough to run at low complexity and medium quality.

• The difference — about $7,600, or roughly the cost of a part-time illustrator for a month.

Now add a second pass. If 15% of the cheap model's images need a human fix and only 5% of the expensive model's do, the arithmetic changes but rarely reverses: at a $4 human-minute, the cheap run still wins. That is the honest shape of this decision. HY Image 3.5 does not win because it looks better. It wins because at catalogue scale, "good enough" is a business model and "best" is a luxury.

Where it flips is when the image is the product: a hero asset, a campaign key visual, a brand illustration. At ten images a week, the price difference is forty dollars and irrelevant. At that volume, buy the better model and stop thinking about it.

Running both

The sensible architecture for most teams is not a choice. Use HY Image 3.5 for volume — drafts, thumbnails, variants, localisation — and Images 2.6 for the assets that get approved and shipped. The two pricing models make that split almost mechanical: a per-image price rewards you for sending cheap work to the cheap model, and a token price rewards you for keeping expensive, high-effort renders rare.

Neither model is on OrcaRouter today. What our image line does carry is GPT-Image-2.5 Flare and Sunburst, Google's Gemini and Imagen family, and Grok Imagine Image — one key, provider list price with 0% markup, automatic failover when a provider degrades. For a two-model split like the one above, that matters less for the model choice than for the plumbing: one integration, one billing line, and a routing rule that sends the cheap job to the cheap endpoint without a second contract.

What to watch: Tencent shipping HY Image 3.5 out of preview, and the first independent Images 2.6 score. The first tells you whether the $0.024 price is a launch promotion or a business model. The second tells you how much of the fourfold gap is quality and how much is just the logo on the invoice.