Hy Image3.5 vs Qwen Image 3.0: The Price Gap Is Three Fen

Set Tencent Hunyuan Hy Image3.5 preview next to Qwen Image 3.0 on a Chinese cloud rate card and the difference is ¥0.15 against ¥0.18 per 2K image. Three fen. About four tenths of a US cent. Both models previewed or launched within seven weeks of each other in 2026, both are 2K-class generators with multi-turn editing and reference-image conditioning, and both come from Chinese labs that price aggressively and publish their own benchmark tables. When two products cost the same and make the same claims, the price column stops being a decision tool and the only thing left that matters is which of them anyone has actually measured.

Qwen Image 3.0 launched on Alibaba's Qwen AI platform on 5 August 2026 with a Standard and a Pro tier and APIs available at the same time, and it went overseas through Qwen Cloud. Hy Image3.5 preview went live on 22 September 2026. Between them they cover almost the entire cheap end of the market, and neither one has been placed head-to-head against the other by anybody outside the two companies.

The rate cards, line by line

Hy Image3.5 is Tencent Hunyuan's native multimodal image model, served as hy-image-v3.5-preview. Text-to-image and image-to-image in one model, multi-turn editing that carries earlier turns as context, up to five reference images per request, standard aspect ratios, 2K output ceiling. Tencent Cloud lists ¥0.15 per 2K image domestically and Tencent Cloud International lists $0.024 per image. You are billed for delivered output images only; discarded and failed generations cost nothing. Preview access covers Yuanbao, ima, Miora, WorkRally, OnSolo and WorkBuddy plus Tencent Cloud TokenHub, MPS and VOD, with a free trial on Miora, WorkRally and OnSolo running 22 September to 7 October.

Qwen Image 3.0 comes in two tiers. Standard is RMB 0.18 per image at both 1K and 2K, with image input at RMB 0.02 and editing input at RMB 0.02 per image. Pro is RMB 0.25 at 1K and RMB 0.50 at 2K. On QwenCloud internationally the Standard tier runs $0.003 input and $0.03 output per image. On Alibaba Cloud Model Studio in USD, the numbers are $0.00275 per 1K input, $0.03438 per 1K generation, $0.068761 per 2K generation, $0.04 per 1K output and $0.075 per 2K output. Throughput limits are published: 10 concurrent requests, a 200-request async queue and 20 requests per minute. Alibaba reported the model at number one in China on Arena.ai's text-to-image leaderboard at launch.

• Price per 2K image, domestic — Hy Image3.5 ¥0.15; Qwen Image 3.0 Standard ¥0.18

• Price per 2K image, Pro tier — Hy Image3.5 has no tiering; Qwen Image 3.0 Pro is ¥0.50 at 2K

• Billing basis — Hy Image3.5 charges delivered output images only; Qwen Image 3.0 also meters image input and editing input

• Reference images — Hy Image3.5 up to five per request; Qwen Image 3.0 conditions on image input at RMB 0.02 per image

• Multi-turn editing — both support it, and Qwen charges separately for the editing input

• Independent placement — Qwen-Image-3.0 Elo 1,077 at #16 and Qwen-Image-3.0-Pro 1,088 at #14 on the Artificial Analysis text-to-image board, from 5,816 and 6,077 samples; Hy Image3.5 has no public arena entry

• Rate limits — Qwen Image 3.0 publishes 10 concurrent, 200 queued, 20 RPM; Tencent has published no throughput limits for the preview

What three fen per image is worth at volume

Run the arithmetic honestly, because it is the only place either vendor's pricing becomes an argument. At ten thousand 2K images a month, Hy Image3.5 costs about ¥1,500 and Qwen Image 3.0 Standard about ¥1,800 — a difference of ¥300, or roughly forty US dollars. At a hundred thousand images a month it is ¥15,000 against ¥18,000, and the ¥3,000 gap is real but it is not the kind of number that decides an architecture.

The number that does decide an architecture is what happens to the attempts you throw away. Hy Image3.5 does not bill them. Qwen Image 3.0 bills image input and editing input separately from generation, so a workflow that feeds a reference image back in for a second pass, then a third, pays at every step regardless of whether the output is usable. If your pipeline iterates three times per accepted image, the effective cost per shipped image diverges fast — and it diverges in Tencent's favour in a way that the headline ¥0.15 against ¥0.18 completely hides. That is the real difference between these two rate cards, and it is a billing-model difference, not a price difference.

Where the evidence is, and where it is not

Artificial Analysis has measured both Qwen Image 3.0 tiers. The Standard model sits at Elo 1,077, sixteenth on the text-to-image board, from 5,816 samples collected in July 2026 at a derived $30 per thousand images. The Pro tier is at 1,088, fourteenth, from 6,077 samples at $40 per thousand. On the editing board the order flips and both drop: Pro at 1,076 and sixteenth, Standard at 1,048 and twenty-first. Those are modest, well-sampled results — mid-table, not frontier, and Qwen's own claim of a number-one position in China on a different leaderboard is a different board with a different population.

Hy Image3.5 has no equivalent. Tencent's published evidence is a blind GSB preference test run with hundreds of professional designers inside the company, reporting roughly a 30% improvement over Hy Image3.0, parity with Seedream 5.0 Pro, and a marginal edge over Nano-Banana Pro and Qwen-Image-3.0 Pro. Vendor-run, vendor-populated, prompt set unpublished, unreproduced by anyone. It is also worth reading precisely: the claim is a slight edge over Qwen-Image-3.0 Pro, which Artificial Analysis places at 1,088 — and there is no claim of any kind against the model this article compares it to on price, the Standard tier.

Two models we do not serve, and one endpoint that matters anyway

Neither of these is on OrcaRouter. We host no Tencent image model — the only Tencent entry in the catalogue is the text-only Hy3 line — so Hy Image3.5 means Tencent's own cloud and several third-party platforms; we host no Alibaba image model either, so Qwen Image 3.0 means Alibaba's own platform and several third-party platforms. Stating that plainly is more useful than a model page that leaves it ambiguous, and it is the reason this comparison has to be made on rate cards and published evaluations rather than on a routing table.

The models we do serve cover the same job from the other direction: openai/gpt-image-2, google/gemini-3.1-flash-image-preview, the Imagen 4 family and grok/grok-imagine-image, all behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price cut is live on our side the same day it happens. Model fusion and the routing DSL let you pin a specific model per request rather than maintaining an SDK per vendor, and automatic failover keeps a degraded route from becoming an outage. If you are evaluating a ¥0.15 image model against a ¥0.18 image model, the useful thing an aggregator provides is not access to either of them — it is a stable, priced baseline to measure both against.

How to call it

If you need published throughput limits, a Pro tier with more headroom, an international API with a documented per-image price, and at least one independent leaderboard placing the model mid-table with six thousand votes behind it, Qwen Image 3.0 is the safer pick and it is not close on evidence. You pay for the input tokens you send, and you pay again when you edit.

If your workload iterates — generate, look, feed it back, generate again — Hy Image3.5's billing model is the better fit at almost exactly the same headline price, and five reference images per request at 2K for $0.024 with no charge for discarded attempts is a genuinely strong offer. What you accept in exchange is that every performance number attached to it comes from Tencent, that the model has no independent placement at all, and that the one comparison Tencent does make is against the Pro tier of the model you are pricing against, not the Standard tier. Run your own prompts through both during the free trial window that closes on 7 October, and let your accept rate decide it.

OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price change is live the same day.