Ideogram 4.5 vs HY Image 3.5: Nine Times the Price, Two Places Behind

On the Artificial Analysis image editing board, HY Image 3.5 sits at 1,087 Elo and $24.00 per thousand images. Ideogram 4.5 sits at 1,064 Elo and $220.00 per thousand. T​encent's preview is ahead by twenty-three points on more than twice the sample, at a ninth of the price. That is the entire comparison, and everything else in this article is a footnote to it — because the footnote is where the decision actually lives. These two models were built for different jobs, sold to different buyers, and priced on completely different theories of what an image costs.

Two previews, five weeks apart

HY Image 3.5 opened on September 22, 2026 as a preview served through T​encent Cloud under the identifier hy-image-v3.5-preview, reachable through the company's TokenHub API surface. It does text-to-image and image-to-image, supports multi-turn editing and multi-reference input, and T​encent's own post caps it at 2K.

Ideogram 4.5 went live eight days later, on September 30, 2026, across Ideogram's app, its own API, its MCP integration and its apps collection, plus partner platforms. Where T​encent leads with price, Ideogram leads with a fidelity claim: that repeated localised edits stop degrading, that the edges of an edited crop are preserved, and that the whole thing works on a 4,016 × 6,016 source without downscaling.

Both are, in the strict sense, previews or near-previews.

Both are also callable and priced today, which is more than can be said for most of what competed with them this autumn.

The spec sheet, side by side

• Price — $0.024 per image on the international endpoint (¥0.15 per 2K image on T​encent Cloud) vs $0.03 at Low, $0.06 at Medium and $0.22 at High.

• Billing trigger — HY Image 3.5 charges only on images returned; Ideogram charges per generation and per edit, at the tier you selected.

• Output — up to 2K on T​encent's own wording vs native 2K with edits demonstrated at 24.2 megapixels.

• Operations — text-to-image, image-to-image and multi-turn editing on both; multi-reference input up to five images on HY Image 3.5, which is the figure attributable to the vendor's documentation.

• Editing score — 1,087 Elo from 5,598 samples, rank 15, at $24.00 per 1,000 images vs 1,064 Elo from 2,459 samples, rank 23, at $220.00 per 1,000.

• Text-to-image score — 1,013 Elo from 2,278 samples at rank 34 of the board for Ideogram 4.5; no text-to-image row for HY Image 3.5 appeared on the board view read for this piece, so the editing row above is the only independent reading available for it.

The nine-times question

The price column deserves a closer look, because it is the one place where the two vendors agree with an independent source. T​encent's published international rate is $0.024 per image. Multiply by a thousand and you get $24.00 — which is exactly what the editing board's own price column reports for HunyuanImage 3.5 (Preview). An independent conversion and a vendor rate card landing on the same number is a good sign that both are being read correctly.

Ideogram's $220.00 per thousand is the board's conversion of the High tier: $0.22 per image. The gap is $196 per thousand images — a factor of 9.2.

At scale that stops being a rounding difference. A catalogue refresh of 5,000 product images costs $120 on HY Image 3.5 and $1,100 on Ideogram 4.5 at High. A monthly content operation pushing 20,000 assets is $480 against $4,400.

The cheaper model is not cheap because it is worse at the thing the board measures — it is, marginally, better at it. It is cheaper because T​encent is pricing for volume distribution in a market where Ideogram is pricing for a specialist workflow.

Which is the actual answer to the headline. If your workload is high-volume image production where "good enough, reviewed by a human" is the standard, the nine-times difference is not a consideration, it is the decision. You are buying throughput, and the throughput model wins.

Where the extra money is supposed to go

Ideogram's case rests on a property neither board measures. Editing arenas show a voter a source image and two results from a single instruction. They do not show a voter the tenth edit in a sequence against the first, which is the failure mode Ideogram built 4.5 to fix: faces that drift a few millimetres across passes, grain and lighting that shift, type that is crisp on pass one and soft on pass four.

If your work is ten small edits on one large file — colourways, signage swaps, restoring a photograph without repainting what survived — that fidelity is worth a multiple. The 24.2-megapixel demo is the vendor's evidence that the edit boundary holds at print resolution, and it is a demonstration rather than a measurement.

T​encent's own quality evidence is weaker on its face. The company says an internal blind human evaluation put HY Image 3.5 thirty percent ahead of Hy Image3.0. That claim is vendor-reported, was run inside the company, and has not been reproduced anywhere public — but it is at least corroborated in direction by the external board, where HunyuanImage 3.5 (Preview) at 1,087 sits well clear of HunyuanImage 3.0 (T​encent Cloud) at 1,029 on the same capture. The vendor's number is unreproduced; the board's is not.

Choosing between them

Take HY Image 3.5 if you are producing images in volume, if your reviewers are human and your standard is "ship it", or if you are building for a market where T​encent Cloud is the natural host and ¥0.15 per 2K image is the rate your finance team already expects.

Take Ideogram 4.5 if the expensive part of your workflow is not generating images but repairing approved ones, and if a lost edge or a drifted face costs more than twenty-two cents. That is a real category of work, and no other model on either board is currently making the claim with a 24-megapixel demonstration attached.

Take neither on the strength of a launch-week number. HY Image 3.5's row carries 5,598 votes and Ideogram's 2,459, on a board where the leaders have twelve to eighteen thousand; both readings are provisional and both will move.

Running the test through one key

Neither of these models is on OrcaRouter's catalogue — HY Image 3.5 is served through T​encent's own cloud, Ideogram 4.5 through the vendor's own API and several third-party platforms, and we route neither. What we do serve from this corner of the market is the G​oogle image line, including the Gemini 3.1 Flash Image identifiers and the Imagen 4 family, plus O​penAI's GPT-Image models.

Which matters here for a specific reason: a nine-times price gap is exactly the kind of claim you should not take from an article, including this one. The test is twenty of your own images, the same instructions on both models, and a reviewer who does not know which output came from which. A routing layer is what makes that test cheap to run — one key, per-call pricing at provider list price with nothing added, and automatic failover so a slow endpoint does not stall the run. If the cheap model holds up on your images, you have just cut your image bill by 89 percent, and you found out in an afternoon rather than a quarter.