Hy Image 3.5 Preview price: $1.60 per million tokens, not per image

Tencent Hy Image 3.5 Preview is listed at $1.60 per million tokens on its OpenRouter page, which is a token rate and not a price per image. You cannot turn it into a per-image cost without knowing how many tokens one of your generations uses, and the page does not give that number.

This matters for anyone budgeting a batch this week, because the listing shows a release on Oct 5, 2026 and most price comparisons you will find quote a per-image figure for other models. Mixing a token rate with a per-image rate gives a wrong total.

What the vendor page states

The OpenRouter listing, read today, names a dated model id, tencent/hy-image-v3.5-preview-20260922, and describes text-to-image, image-to-image and multi-turn editing, up to 20 reference images and output up to 4K. The price line reads $1.60 per million tokens.

The page does not state how many tokens a 1024x1024 or a 4K output consumes, so any per-image dollar figure for Hy would be a guess. This post does not give one.

How Sume prices an image instead

Sume meters image models per image, not per token. The Image API returns a usage.cost field in USD, and the token counts in usage are always 0 in v1. A failed or cancelled generation is not charged, as the Image API docs describe.

Sume does not list a Hy model today. You can confirm that for your own workspace by calling GET /v1/images/models, which returns the models with their capabilities. If Hy appears there later, the catalog row, not this post, is the price to trust.

• Token-priced model: cost = tokens used x rate, so it varies with prompt, references and size.
• Per-image model on Sume: the response reports the cost in usage.cost.
• Preview models can change price, so re-read the page on the day you budget.

Worked arithmetic, with a placeholder count

The formula is simple: cost = tokens used divided by 1,000,000, times $1.60. As arithmetic only, a hypothetical job that used 1,000 tokens would cost $0.0016, and one that used 10,000 tokens would cost $0.016. Those token counts are placeholders chosen to show the formula, not measurements of Hy Image 3.5 Preview, and the vendor page gives no figure to replace them with.

The reason to run the formula on your own numbers is that the spread between a small and a large job can be wide, and a flat per-image guess hides it. Record tokens per job for a week, then use the median and the 90th percentile for your forecast.

• Log the token count the vendor returns with each result.
• Log size, number of references and whether the call was a first turn or an edit turn.
• Compare the median with the worst case before you promise a client a flat price.

How to budget a token-priced image model

Run a small sample of real jobs, record the tokens the vendor reports for each, and multiply by the rate. Do this at each size and with and without references, since reference images are inputs and may count too. The vendor page does not say whether they do, so measure it instead of assuming.

If you need a price you can fix in a spreadsheet today, use a model that already has a per-image row. For a Sume example of a fully calculated grid, see the GPT Image 2.5 xhigh and max price grid.

A last point on comparing quotes: when two vendors quote in different units, convert both to cost per approved image, not per generated image. If a model needs three tries to reach an acceptable result, its effective price is three times the sticker price. Sume's billing note matters here: a failed or cancelled generation is not charged, but a completed one that you reject still is. Track your own acceptance rate for any model you test, preview or not, and divide spend by approved outputs. That single number is more useful in a budget meeting than any list price, and it works for token-priced and per-image-priced models alike.