A generated title card reading 'Kling Image 4.0 vs HY-Image 3.5' with the subtitle 'A 208-point gap, a 91-point one'. Two horizontal bars run left to right: the one labelled text-to-image is split into a short grey segment marked 208 and a longer blue one, and the one labelled image editing shows a shorter split marked 91. Below, a price row reads 6.1 cents against 2.4 cents per 2K image.
Guides & Insights

Kling Image 4.0 vs HY-Image 3.5: A 208-Point Gap, a 91-Point One

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

The price ratio between Kling Image 4.0 and HY-Image 3.5 is 2.6 to one. The quality ratio is not one number at all: Artificial Analysis puts 208 Elo between them when both generate a picture from scratch, and 91 Elo between them when both are asked to edit one that already exists. Same pair of models, same board, same month of measurement, two answers that differ by more than a factor of two. Anyone quoting a single score for this matchup is quoting half of it.

The dates matter to the frame. Kling Image 4.0 has been live since 17 November 2026, sold on Kuaishou's unit ladder at 16 units for a 2K still — about 6.1 cents at the entry package rate, or $63 per thousand on the Artificial Analysis listing. HY-Image 3.5 is older and hangs on a different word: Tencent put it out on 22 September 2026 as hy-image-v3.5-preview, at ¥0.15 or US$0.024 per 2K image, and the independent boards have only recently started scoring it. Its 972 Elo is not a launch claim. It is the first outside measurement of the model, and it lands below Tencent's own predecessor.

A screenshot of Kling's release-notes page in English, headed Release Notes, with the newest entry dated 2026-11-17 marked Major and titled 'Kling Image 4.0 Is Here!', described as stunningly realistic and sharp with results from a single image to a cohesive series, above an older 2026-10-15 entry for the Kling 4.0 video model and a 2026-06-17 entry for the Kling 3.0 series.

The two gaps, and why they differ

Both models are scored on the same two Artificial Analysis boards, which is what makes the split readable rather than rhetorical. Read against the field on 18 November 2026:

• Text-to-image Elo — Kling Image 4.0 1,179.93, seventh of 115; HY-Image 3.5 972.12, seventy-fifth of 115

• Image-editing Elo — Kling Image 4.0 1,182.89, sixth of 87; HY-Image 3.5 1,091.87, twenty-eighth of 87

• Confidence intervals — Kling 1,166.93 to 1,192.93 against HY 960.12 to 984.12 in generation, and 1,168.89 to 1,196.89 against 1,078.87 to 1,104.87 in editing

• List price — Kling $63.00 per thousand images in generation and $66.00 in editing; HY $24.00 on both boards

• Published unit rate — Kling 16 units per 2K still, 24 at 4K, 40 at 8K, at roughly $0.00384 a unit; HY $0.024 per 2K image, billed on delivered output only

• Output ceiling — Kling documents a native 8K tier; HY documents 2K as the ceiling

• Reference inputs — Kling bills 0.8 units per reference image or reference element, newly chargeable in 4.0; HY accepts up to five reference images per request and publishes no separate reference fee

• Weights and identifier — both proprietary; HY has no general-availability identifier and no open checkpoint

Neither interval overlaps, so both orderings survive the obvious objection that the numbers are noisy. The interesting part is the size. A 208-point gap on a 400-point Elo scale implies the stronger model wins roughly three pairwise comparisons in four. A 91-point gap implies about three in five. Both are real advantages; only one of them predicts a visible difference on your own prompt set most of the time.

The shape of that split is a statement about what each model is for. HY-Image 3.5 sits in the bottom third of a 115-model generation board and in the top third of an 87-model editing board. That is not the profile of a model that is broadly weaker. It is the profile of a model tuned for the second task, and it should change who buys it: if your work is editing an existing frame — a product photo, a character sheet, a client's layout — the cheaper model is 91 points behind, not 208.

Tencent's own predecessor scores higher

The most inconvenient number for Tencent in this pairing is not Kling's 1,179. It is 991.90: the score Artificial Analysis gives HunyuanImage 3.0 (Tencent Cloud), the previous generation of the same line. HY-Image 3.5 preview scores 972.12 on the same board. The newer model is 20 Elo below the older one at text-to-image, with a confidence interval that clears the gap.

What Tencent claimed at launch was the opposite direction. Its evidence was a blind preference study run inside the company with several hundred of its own designers, reporting roughly a 30% gain over Hy Image 3.0, parity with Seedream 5.0 Pro and a marginal edge over Nano-Banana Pro and Qwen-Image-3.0 Pro. That study has never been published as a dataset anyone can re-run, and the independent board now disagrees with it on generation.

The disagreement is narrower than it first looks, because the editing board tells the other half of the story. On image editing, HY-Image 3.5 scores 1,091.87 against 1,028.30 for HunyuanImage 3.0 (Tencent Cloud) and 1,063.62 for the Instruct variant — a 64-point gain over the better of the two. So the improvement Tencent measured is real; it just lives in editing, and the generation half of the claim did not travel. One caveat belongs on the record: the board's Hunyuan rows are separate listings, and a listing change can move a score for reasons that have nothing to do with weights. Treat the 20-point regression as a flag pointing at where to test, not as a verdict on the checkpoint.

What the 2.6x buys on Kling's side

Kling's answer to "why 2.6 times the price" is not the leaderboard, where it sits seventh. It is the format ladder. HY-Image 3.5 stops at 2K. Kling Image 4.0 sells 4K at 24 units and 8K at 40, and it sells two output shapes Tencent does not document at all: a coherent set from a single prompt, and long, page-shaped images — 140 units for a webpage comp at 1K and 152 at 2K, or an e-commerce detail page at 60, 120 or 200 units depending on length, image-to-image only.

If your deliverable is a 900-pixel product thumbnail, that ladder is a spec sheet for features you will never use, and paying 6.1 cents instead of 2.4 is paying for unused ceiling. If it is a print asset, a full-page layout or a twelve-frame campaign set that has to look like one shoot, the comparison ends before quality is discussed: one model sells the tier, the other does not have it. That is a capability difference, and no Elo gap captures it.

The reference bill cuts the other way and is easy to miss. Kling Image 4.0 is the first model in its line to charge for reference images and reference elements, at 0.8 units each on top of the output price. A reference-heavy set built from four inputs moves from 192 units of output to 230 units of total — about a fifth more, scaling with how reference-dependent the work is. HY-Image 3.5's rate card charges delivered output images only and publishes no reference surcharge. On the most reference-heavy workflows the 2.6x price ratio is not the real ratio.

A generated two-column comparison scoreboard titled 'Kling Image 4.0 vs HY-Image 3.5 — the scoreboard'. The Kling column reads: text-to-image Elo 1,180, rank 7 of 115, image-editing Elo 1,183, rank 6 of 87, price $63 per 1,000 images, ceiling native 8K tier, references billed 0.8 units each, sets and long-page modes yes. The HY-Image column reads: text-to-image Elo 972, rank 75 of 115, image-editing Elo 1,092, rank 28 of 87, price $24 per 1,000 images, ceiling 2K, references up to five per request with no separate fee, sets and long-page modes none documented. A footer reads 'Scores per Artificial Analysis, read 18 November 2026. Prices per each vendor rate card.'

The preview label is a procurement risk

There is a second difference in this matchup that has nothing to do with pixels, and it is the one a finance or legal review will raise. Kling Image 4.0 is a shipped model with a release-notes entry, a published unit price and a version number that implies a successor will be numbered, not renamed. HY-Image 3.5 exists only as hy-image-v3.5-preview. Tencent has published no general-availability identifier, no date for one, no open checkpoint and no offline route; every documented way to reach the model is Tencent Cloud's TokenHub surface or one of Tencent's own applications.

That is not a reason to avoid the model, and the price is the reason not to. It is a reason to scope the integration: anything pinned to a preview identifier should sit behind an interface you control, because the name you write into a config file is the one thing about this model Tencent has not committed to. Kling's equivalent risk is a price revision, which at least shows up on a rate card before it shows up on an invoice.

Running the bake-off on one key

To be precise about our own catalogue before the recommendation: OrcaRouter serves neither Kling Image 4.0 nor HY-Image 3.5. Any test of this pair means Kling's API and Tencent Cloud respectively, and the prices above are the vendors' own, not quotes from us. What OrcaRouter does change is the shape of the experiment around them. Image models we do serve sit behind one OpenAI-compatible endpoint at the provider's list price through at 0% markup, so the rate you budget is the rate the vendor publishes and a vendor price cut is live the same day rather than at the next repricing cycle. With more than 200 models on one key, the incumbent lane and the challenger can be called from the same code path, and automatic failover keeps a batch moving when one route refuses a request mid-run.

That matters here because the honest test is an invoice comparison, not a picture comparison. Twenty prompts through both models at 2K, with the same three reference images attached, will settle three things faster than any launch thread: whether the 208-point generation gap is visible on your subject matter, whether the 91-point editing gap is, and what the reference surcharge does to the arithmetic on the set you actually build.

A screenshot of the OrcaRouter pricing page in English, headed 'We never mark up the tokens you buy', with a Zero Markup Guarantee panel stating that routing is free and that the rate paid is the provider's rate, and a rate table listing models with input, cached input, cache write and output prices per million tokens beside an OrcaRouter fee column reading $0.00.

What to test before you commit

Start with the job, not the score. If the deliverable is an edit — a retouched product shot, a character held consistent across a run, a client's layout with one element changed — HY-Image 3.5 is the value play in this pair and the 91-point gap is small enough that your own reference images may close it. It is a quarter of the price, it bills only what it delivers, and it does not charge for the references the work depends on.

If the deliverable is a still that has to be large, or a set that has to look like one shoot, the generation gap is the one that governs, and 208 points is not a margin you tune away with prompt engineering. Kling Image 4.0 is the more expensive instrument and the only one in this pair that sells the tier you need.

The one reading to discard is the vendor's. Tencent's 30% claim points at a board where its own newer model now measures 20 points below its predecessor on generation and 64 above it on editing, and any plan built on the average of those two numbers is built on a number nobody published.