Ideogram 4.5 vs Qwen Image 3.0: Nine Points of Editing, Seven Times the Price

On Artificial Analysis's image-editing leaderboard, Ideogram 4.5 sits at an Elo of 1,064 and Q​wen Image 3.0 sits at 1,055. Nine points. On price, Ideogram charges $0.22 for a High-resolution edit with a source image and Model Studio charges $0.035 for a generation. Just over six times, or seven and a third against the $0.03 the measurement board derives for the vendor's Standard tier, for nine points of Elo that will not sit still, on a model that is one day old. But that is the comparison only if you insist on firing Ideogram's most expensive tier at the vendor's only one, and Ideogram's own ladder runs the other way as well: at its Very Low setting the same model costs $0.008, roughly a quarter of that price. Everything worth saying afterward is about which end of that ladder your workload actually sits on — and about a question neither vendor's launch page answers clearly, which is what happens to the open-weights story each of them is trading on.

Two numbers, read honestly

Start with the caveat, because it carries the rest of the piece. Ideogram 4.5's 1,064 comes from 2,382 votes collected in the day since launch; the 95% interval runs from 1,053 to 1,075. Q​wen Image 3.0's 1,055 comes from 7,594 votes and its interval runs from 1,047 to 1,063. Those overlap. The nine-point gap is currently indistinguishable from noise, and Ideogram's number will move as votes accumulate. Anyone quoting 1,064 > 1,055 as a settled result is over-reading a one-day sample.

The text-to-image board is a different picture and it does not flatter Ideogram at all: Q​wen Image 3.0 is sixteenth at 1,075 and Qwen-Image-3.0-Pro fourteenth at 1,089, against Ideogram 4.5 (High) thirty-fourth at 1,014. That is a sixty-point gap in A​libaba's favour, from sample sizes between two-and-a-half and three times larger, and the interval on Ideogram's number does not come close to covering it.

So the honest shape before any pricing is this: Ideogram 4.5 is plausibly the better editor by a margin too small to measure, and measurably the weaker generator by a margin that is not close.

What Ideogram 4.5 is selling, and what it costs

Ideogram 4.5 is a precise-edit model released on 30 September 2026. Its two claims are drift reduction across multi-turn edits — "Ideogram 4.5 reduces this drift, preserving details across multi-turn edits" — and in-place editing at high resolution, demonstrated on a 4,016 × 6,016 pixel, 24.2-megapixel source where the model preserves the edges of an edited crop so it can be stitched back into the original without downsampling anything else.

Ideogram prices it per generated image, by rendering speed and output size. Very Low is $0.008, Low $0.03, Medium $0.06, High $0.10, and a High request that includes a source image — the image-to-image path — is $0.22. The Precise Edit endpoint mirrors those numbers exactly. Output size from 1K to 2K changes the price by nothing at all, which is an unusual and deliberate piece of pricing: Ideogram is charging for compute per render, not for pixels delivered.

What Q​wen Image 3.0 is selling

Q​wen Image 3.0 is A​libaba's image model, shipped in two tiers. Artificial Analysis records its release as 21 July 2026; A​libaba's own launch material put the Standard and Pro tiers live on Model Studio on 5 August 2026, and we flagged the same two dates when we covered the T​encent Hunyuan comparison in September. It is a general-purpose generator and editor, priced far flatter than Ideogram's: A​libaba's Model Studio documentation lists image generation at $0.035 per image for qwen-image-3.0, and Artificial Analysis's derived figures are $30 per thousand for the Standard model and $40 per thousand for Pro. The domestic RMB rate card we read earlier this year was ¥0.18 per image for Standard and ¥0.50 at the Pro tier's 2K setting, with image input and editing input metered separately at ¥0.02 each — so a workflow that feeds a reference image back for a second pass pays again on the way in. There is no low-cost draft tier and no high-cost precision tier. One price, or two if you count Pro.

• Price per generation — Ideogram 4.5 $0.008 to $0.10 by rendering speed; Q​wen Image 3.0 $0.035, or about $0.03 per the derived board figure

• Price for an edit with a source image — Ideogram 4.5 $0.22 at High, $0.03 at Low; Q​wen Image 3.0 output price plus separately metered image input and editing input

• Published price tiers — four at Ideogram, one at Q​wen plus a Pro variant

• Independent editing rank — Ideogram 4.5 (High) twenty-third at Elo 1,064 from 2,382 samples; Q​wen Image 3.0 twenty-fifth at 1,055 from 7,594 samples

• Independent text-to-image rank — Q​wen Image 3.0 sixteenth at 1,075 and Qwen-Image-3.0-Pro fourteenth at 1,089; Ideogram 4.5 (High) thirty-fourth at 1,014

• Searchable reference — Q​wen Image 3.0 can search the web while generating, which is a generalist feature Ideogram does not advertise

• Demonstrated edit resolution — Ideogram 4.5 shows a 24.2-megapixel source edited in place; Q​wen Image 3.0's published ceiling is a 2K output tier

The open-weights question, which neither launch page answers cleanly

This is where the two vendors' stories are more tangled than their marketing suggests, and it matters because both are competing for the same developer who reads licences.

Q​wen Image 3.0 is not open weights. A​libaba's open image line stops at Qwen-Image-2.1, which sits on the boards with a Hugging Face link and no commercial API at all. Q​wen Image 3.0 appears only as a metered API product on Model Studio, and Artificial Analysis lists it without an open-weights link. If you adopted Q​wen's image line because the weights were downloadable, the current generation is the point at which that stopped being true.

Ideogram's situation is the mirror image and is, if anything, murkier. Ideogram 4.0 is genuinely open — 9.3 billion parameters, nf4 and fp8 checkpoints on the company's own Hugging Face organisation under a non-commercial licence, a public inference repository, and the GitHub and Hugging Face links sitting in the main site navigation today. Third-party launch coverage of 4.5 said Ideogram "is planning to publish open weights of this model soon as well." Ideogram's own 4.5 page does not say that. The page describes an edit model, shows its examples, and mentions no licence, no repository and no release plan; the linked repositories and the collection still contain the 4.0 line only. So the correct state of that claim is: reported, not committed — a third-party statement about a vendor's intentions, unrepeated by the vendor, and therefore not something to plan an on-premise deployment around.

Put the two together and the open-weights scoreboard for this matchup reads: Ideogram's previous generation, open and non-commercial; Ideogram's current generation, closed with a rumoured opening; Q​wen's previous generation, open; Q​wen's current generation, closed. Nobody shipping an image model in this generation has handed over the weights yet, and the one company with a demonstrable track record of doing so has not said it will.

Working out the multiplier

Take a realistic pipeline rather than the list price, because the list price hides the decision. Suppose you accept a Low render most of the time and escalate to High with a source image on the one in twenty jobs that needs real precision. Your blended Ideogram 4.5 cost per image is somewhere just above $0.03; your Q​wen Image 3.0 cost is $0.035 plus whatever the editing input charges add on the retries. On that mix the two are close enough that the difference disappears into your accept rate — which is the honest conclusion, and it is the opposite of what comparing $0.22 against $0.035 suggests.

The ladder is the point. Ideogram's tiers are rendering speeds, not quality classes; the model is the same model at every price. What you are actually buying at $0.22 is the right to spend eight minutes of GPU time on one image and to edit it in place at 24 megapixels. If your work generates a lot of images and keeps few, the Very Low and Low tiers make Ideogram the cheaper of the two by a wide margin. If your work edits a small number of large assets, Q​wen Image 3.0 cannot do the job at all — its published output ceiling is 2K — and the price comparison is beside the point.

Routing, and what is actually on the catalogue

Neither of these image models is on OrcaRouter. We carry A​libaba's Q​wen text line — Qwen3.8 Max and its siblings — but no Q​wen image model, and no Ideogram model on any tier, so Ideogram 4.5 means Ideogram's own app and API and Q​wen Image 3.0 means A​libaba's Model Studio. Third-party platforms host both; we are not one of them.

What we are is the stable side of the comparison. OrcaRouter runs 200+ models behind one OpenAI-compatible endpoint with provider list price passed through at 0% markup, which in this matchup is the difference between one key and two vendor accounts, two SDKs and two billing relationships. Model fusion is the part that applies here specifically: a panel of models answering the same request is a cheap way to find out whether an image task has a confident answer at all, before you commit to a precision tier at $0.22 a render. And automatic failover means the baseline you measure against does not move because a single provider had a bad afternoon mid-benchmark — which matters a great deal when the number you are trying to establish is a nine-point Elo difference you cannot yet resolve.

Which one to run

If you are generating — filling a catalogue, producing social variants, making images at a known size for a known pipeline — Q​wen Image 3.0 is the better buy on the evidence available today. It is measured sixty points higher than Ideogram 4.5 on the board with the most votes and the tightest intervals, it is roughly a third of the price of Ideogram's High generation tier, it costs the same whether you ask for 1K or 2K relative to A​libaba's metering, and it grounds on the web. Its weakness is billing rather than quality: input images and editing passes are metered separately, so an iterative workflow pays at every step whether or not the output is usable.

If you are editing — one large existing asset, changed in one place, kept at full resolution — Ideogram 4.5 is the only one of the two claiming the job, and the $0.22 tier is the price of the claim rather than a margin. The reason to hesitate is not the price. It is that the property you are actually paying for, stability across a sequence of edits, has no independent measurement anywhere, and that the one-day editing sample putting 4.5 nine points above Q​wen Image 3.0 has error bars wide enough to swallow the difference whole.

Wait on the licence question in both directions. A​libaba has already closed a line that used to be open, and Ideogram has a genuinely open previous generation with a third-party report — not a vendor commitment — that the current one will follow it. Whichever way that resolves, it changes the calculus for anyone running on-premise far more than nine points of Elo does.