Krea 3 vs HY-Image 3.5: The Cheaper Editor Is the One With the Worse Documentation

On price alone this matchup is over before it starts, and in the wrong direction for the newer model. HY-Image 3.5, T​encent's current image model — served under the identifier hy-image-v3.5-preview and priced at $0.024 for a 2K image on the international endpoint — costs roughly forty per cent of what K​rea 3 charges. K​rea released K​rea 3 on 15 October 2026 as a dedicated editing model at $0.06 per image at native 2K, selling speed: under 8 seconds per edit, end to end. So if you arrived expecting the new challenger to undercut the incumbent, the arithmetic says the opposite, and the interesting question becomes why anyone would pay two and a half times as much for the same editing job.

The answer is legibility, and it cuts both ways. T​encent's model has an independent editing score — 1,088 Elo, seventeenth of sixty-nine ranked models on the Artificial Analysis image-editing board, on a stated 95% interval of 1,074 to 1,102 across 5,775 blind comparisons. K​rea 3 has none. But T​encent's model has been in preview since well before K​rea 3 existed, has no published date for leaving preview, and has no published seconds-per-edit figure at all. K​rea 3 has a release date, a price, a latency number and no external evidence. Neither vendor has given you a complete picture, and the two gaps are not symmetrical.

Two models, two different kinds of unfinished

This is the first thing to get straight, because the word "preview" on one side and the word "beta" on the other are not describing the same problem.

T​encent ships its H​unyuan image line preview first and official version later, and the current generation is in the first of those states — there is no second, stable, non-preview identifier sitting beside it, and T​encent has published no date for when one will appear. The board registers it under the preview designation too, as HunyuanImage 3.5 (Preview), which is independent corroboration that preview is the current state rather than a historical label.

K​rea's situation is messier in a smaller way. The model launched on 15 October 2026 as K​rea 3, and K​rea's launch page and blog call it that. K​rea's own pricing tables list it as K​rea 3 Beta. Elsewhere on the same site the reference-image limit is given as six on the launch post and eight on the product page. None of that suggests the model is unready; all of it suggests a release that is four days old and has not been tidied. But "K​rea 3 Beta" is the string in the plans table, and that is the sort of detail that ends up in somebody's config file.

• Release state — HY-Image 3.5: in preview, identifier hy-image-v3.5-preview, no announced date for a stable designation. K​rea 3: released 15 October 2026, named K​rea 3 on the launch page and K​rea 3 Beta in K​rea's plan tables.

• Independent editing rank — HY-Image 3.5: #17 of 69, 1,088 Elo, interval 1,074–1,102, 5,775 samples. K​rea 3: no entry, unranked.

• Price per 2K image — HY-Image 3.5: $0.024 on the international endpoint, ¥0.15 domestic, billed on delivered output images only. K​rea 3: $0.06 at 2K, $0.03 at 1K, flat, with K​rea stating the launch price is the list price rather than a promotion.

• Speed — HY-Image 3.5: no published seconds-per-edit figure. K​rea 3: under 8 seconds end to end at 2K, vendor-stated.

• Reference images — HY-Image 3.5: up to five per request, the figure carried consistently in T​encent's material. K​rea 3: six on the launch post, eight on the product page.

• Output ceiling — HY-Image 3.5: "up to 2K" in T​encent's own words. K​rea 3: native 2K, with examples quoted at 2736 × 1536 px, and 1K at half price.

• Open weights — neither. No downloadable checkpoint from either vendor; T​encent's ComfyUI nodes call the API rather than running locally, and K​rea offers self-hosting only as an enterprise licence conversation.

What an editing Elo does and does not buy you

Sixteenth of eighty-one is a genuinely good placement, and it is the only independently produced number in this entire comparison. It is worth reading precisely.

HY-Image 3.5 scores higher on the editing board than on the text-to-image board — 1,088 against a text-to-image placement outside the top sixty — and that split is the model's character in one line. On raw generation preference it is mid-field at best. On editing it is top-quarter. The leaders of the editing board are O​penAI's GPT Image 2.5 Sunburst at 1,198 and Flare at 1,177, with Nano Banana 2.1 and M​icrosoft's MAI-Image-2.6 tied at 1,148, so T​encent sits about 110 Elo below the frontier — a real gap, not a rounding difference, on intervals of roughly ±14.

What that number cannot tell you is the thing a production editor lives or dies on. Elo is an average of blind preference votes; it says nothing about whether the unchanged part of your frame stayed unchanged, which is exactly what K​rea has built its entire pitch around with a metric it calls precision — the share of pixels outside the edit area left unchanged, compared at quarter resolution after alignment, one attempt per model at the same prompt. K​rea's precision test is self-run and should be read as such. It is still the right question, and the independent board does not answer it for either model.

The asymmetry nobody advertises: multilingual text

There is a capability difference here that neither vendor's comparison page will surface, and for teams shipping outside English it is the decisive one.

T​encent's claim for HY-Image 3.5 is improved accuracy on text rendered inside the image, in Chinese and English both. That is a vendor claim, no independent board scores this model on text rendering, and it has not been reproduced outside T​encent — but it is a claim about a capability the model is explicitly built for, and it is aimed at the category's most reliable failure. Anyone producing Chinese-language campaign assets, packaging or posters should read it as the only thing either side of this matchup has said on the subject.

K​rea 3's published material says nothing comparable. Its launch examples are in English and its documented strengths are product scene changes, object removal and photo cleanup. That is not evidence of weakness in Chinese; it is an absence of any claim, which for a localisation pipeline amounts to the same thing — untested.

Both models have a shared blind spot in the same direction, incidentally. K​rea 3 keeps character sheets unnamed but real: consistency across a sequence of edits plus references. T​encent documents multi-reference conditioning and a consistency claim, and equally documents no named character-sheet feature or per-character asset object. If you are building a recurring cast, you are assembling the pipeline yourself on either side.

The cost argument, with the retries included

Take the published rates at face value first. HY-Image 3.5 at $0.024 per 2K image against K​rea 3 at $0.06 is a factor of two and a half, and at a hundred thousand images a month that is $2,400 against $6,000. T​encent's billing is on delivered output images only, so a filtered or failed request is not charged, which makes the effective rate closer to the quoted one on messy workloads than a per-request bill would be.

Two caveats have to ride along with the cheaper figure. The $0.024 is the international endpoint rate; the domestic figure is ¥0.15 for a 2K image, and if your traffic is routed differently the number you pay changes. And the model is in preview — a preview price is the price of a thing that has not finished changing, which is an argument for holding the rate as a configuration value rather than a constant, not an argument against using it.

K​rea's counter-argument is not the price, it is the retry count. At six cents an edit, running five passes and picking the best one costs thirty cents and no meeting. At six cents against $0.024, the gap is small enough that the decision should be made on containment and speed rather than on the invoice — which is a strange conclusion for a comparison that opened with a two-and-a-half-times price difference, and it is the honest one.

Where you can call each, and what a router does about it

Neither model is one you can hold in your own hands. T​encent publishes no checkpoint, no local licence and no offline route; every documented way to reach it is T​encent's own API or one of T​encent's own products. K​rea's self-hosting answer is an enterprise plan running on your own GPUs under a negotiated licence. So both are calls, and the practical question is how many calls you want to manage.

OrcaRouter's answer to that is one OpenAI-compatible endpoint over a few hundred models, with provider list prices passed through and nothing added — so a vendor reprice lands on your bill the same day rather than at the next renewal — plus a routing DSL for pinning a specific model per request and automatic failover when a provider errors. Being straight about coverage: we do not route HY-Image 3.5 or K​rea 3 today, so both mean calling the vendor directly. What the endpoint does buy you in the meantime is that the models you would benchmark either of them against — GPT Image 2 and G​oogle's newer image models among them — are already behind one key, so the bake-off is a configuration change rather than three integrations, and a preview-to-stable transition on T​encent's side or a K​rea 3 Beta rename is a one-line edit rather than a migration.

Which one, for which job

If you ship Chinese-language assets, HY-Image 3.5 is the only one of these two with a claim on the table about text rendering, and it is also the cheaper of the two by a wide margin. Go in understanding that preview is its current release state, that a stable identifier has not been dated, and that the only independent quality number attached to it is an average of blind preferences that says nothing about containment.

If you are editing English-language product and campaign imagery at volume and your team's complaint is wall-clock time, K​rea 3 is the newer, faster, better-documented-in-the-specific-sense option — and the more expensive one. Go in understanding that every performance figure attached to it is K​rea's own, that its reference count is documented two ways, and that it entered the world four days ago without a single external measurement.

The tie-breaker is not in either vendor's material. It is whether your edits must hold a frame exactly, in which case the metric to run on your own images is K​rea's precision test on both models, or whether your assets must render Chinese correctly, in which case the test is a headline and not a leaderboard. Both experiments cost less than a day, and both will tell you more than the two price tags that opened this piece.

What the numbers cannot tell you is whether any of it has gone wrong in the field: the documented AI incidents and agent failures around models in this class are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.