Ideogram 4.5 vs Grok Imagine Image 2.0: Same Price Question, Opposite Answers

Here is the comparison in one line: on Artificial Analysis's editing board, Grok Imagine Image 2.0 ranks 10th at 1,107 Elo and is priced at $60 per 1,000 images. Ideogram 4.5 ranks 23rd at 1,064 Elo and is priced at $220 per 1,000 images on the same board's conversion. Cheaper by a factor of 3.7, ahead by 43 Elo. On the text-to-image board the pattern repeats — Grok Imagine Image 2.0 at 1,155 against Ideogram 4.5's 1,014, at $60 against $100.

Both models are editing-first products from labs that decided editing is the frontier. Both shipped this year — Grok on August 7, 2026, Ideogram 4.5 on September 30, 2026. And they answer the same question in opposite ways: xAI built an editing environment with a suite of user-facing tools, Ideogram built an editing engine optimised for one property. Knowing which of those two shapes fits your work is more useful than the Elo gap, but the Elo gap is real and it runs in one direction.

The four numbers that frame it

• Text-to-image — Grok Imagine Image 2.0 4th of 167 at 1,155 ± 11 Elo (6,879 votes, August 2026), $60.00 per 1,000 images. Ideogram 4.5 34th at 1,014 ± 11 Elo (2,247 votes, September 2026), $100.00 per 1,000.

• Image editing — Grok Imagine Image 2.0 10th at 1,107 ± 9 Elo (7,308 votes), $60.00 per 1,000. Ideogram 4.5 23rd at 1,064 ± 11 Elo (2,382 votes), $220.00 per 1,000.

The editing row is the one that decides this matchup, because both models are being sold primarily as editors. A 43-Elo gap with non-overlapping intervals, at a price 3.7 times higher, is not a subtle result. It does not mean Ideogram 4.5 is a bad model. It means that on the only independent measure the market has, the cheaper model produces edits that a blind panel prefers more often.

Two housekeeping notes on those figures. The board lists Grok Imagine Image 2.0's creator as SpaceXAI, which is how xAI appears in its data — that is the board's label, not a mistake. And Ideogram 4.5's rows are two weeks old with roughly a third of Grok's sample size, so its ±11 interval is the widest kind of estimate; the gap is real but the position could move a little as votes accumulate.

An engine and an environment

The scores matter less than the shape of each product, and the shapes are genuinely different.

• Ideogram 4.5 — an engine. Two operations, generate and precise edit, four rendering speeds, native 2K, and a single headline property: edits preserve the boundary, so a crop survives being stitched back into a 4,016 × 6,016 original without re-rendering the rest of the frame. You bring the instruction; it does the pixels.

• Grok Imagine Image 2.0 — an environment. Magic Wand for region-selective edits, segmentation for recolouring or replacing precise areas, background removal, multi-reference input (up to five images in the consumer apps, three in the API), Smart Resize to recompose one image into nine aspect ratios, and templates for product photography, headshots, e-commerce, game assets and marketing posters.

Read those two lists and the pricing stops looking strange. Ideogram is charging for a fidelity guarantee on a single operation. xAI is charging one flat rate for a toolkit. If your work is "apply this exact instruction to this exact region of a 24-megapixel file and keep everything else bit-identical," Ideogram's offer is the more specific one. If your work is "make this product shot work in nine formats and swap the background," the toolset already covers it and costs less.

The claim neither board tests

Ideogram 4.5's specific promise is that repeated edits stop drifting — no pixel shifts, no colour creep, no artifact buildup across turns. Its launch page demonstrates exactly that, and the demo names Grok's sibling generation among other models as the comparison point.

No leaderboard measures this. The editing arena shows voters a single source image, a single instruction and two results. A voter cannot see that pass four quietly softened some type or shifted a face by two millimetres, because pass four is not on screen. Multi-turn fidelity is invisible to the method.

So the fair reading is narrow but real: on the property both vendors actually sell — editing quality as judged one edit at a time — Grok Imagine Image 2.0 is ahead on more evidence and costs three-quarters less. On the property only Ideogram claims — fidelity across a sequence — there is no public measurement for either model, and the only way to get one is to run your own.

That asymmetry is worth naming plainly, because it is easy to read the scoreboard as a settled verdict and easy to read Ideogram's launch page as a rebuttal. Neither is true. The scoreboard measures something adjacent to the claim; the launch page asserts the claim without evidence a third party can check.

What BFL-style multi-turn editing costs you

If drift is your bottleneck, the arithmetic is not close to neutral. Ideogram 4.5 runs $0.03 per image at Low, $0.06 at Medium and $0.22 at High on the reseller rate cards published at launch. The drift demonstration — 24-megapixel source, preserved edit boundary — is a High-setting workload. A ten-turn edit sequence on one image is $2.20 in Ideogram credits before you have decided whether it worked.

Grok Imagine Image 2.0 at $60 per 1,000 is $0.06 per image. Ten turns on the same image is sixty cents, and if the tenth turn has drifted you have lost less than a third of what the Ideogram run cost you to get to the same disappointment — or, more usefully, you can afford to run the experiment on five times as many images to find out whether drift is actually a problem in your pipeline.

That is the practical case against starting with Ideogram for a fidelity question you have not yet measured. Run the cheap experiment first; buy the specialised engine only for the images where the cheap tool demonstrably fails.

The availability wrinkle

Neither model is on OrcaRouter in the generation being compared. Our catalogue carries xAI's earlier image model, grok/grok-imagine-image, at $0.20 per call — that is the previous generation, not the 2.0 release, and it is not the row being discussed above. Ideogram does not appear on our catalogue at all. The comparison here is between two hosted products reached through their own vendors' APIs and, in both cases, through third-party platforms.

Our OpenAI image line is a useful illustration of the opposite situation. GPT-Image-2 sits at $8 per million input tokens and $30 per million output tokens, which is exactly OpenAI's own list price — zero markup, provider rate passed through. When a vendor on the catalogue cuts a price, that cut is live on our side the same day without a renegotiation, which is the property that makes a routing layer worth having for image work. The models in this article are not on that list yet, so the honest statement is simply that they are not routable here today.

If either arrives, the evaluation pattern is unchanged and cheap: point a slice of traffic at the new identifier, keep the incumbent as automatic failover, and let the new model earn its place on real images rather than on a demo reel. That is a materially better way to test a drift claim than a two-image side-by-side.

Which one to pick

Choose Grok Imagine Image 2.0 if you want one flat rate covering generation, editing and the surrounding toolkit, and if your images are the size a browser tab can hold. It is fourth in the world at generation, tenth at editing, cheaper than everything around it on both boards, and its reference at 1,155 Elo from nearly 7,000 votes is a stable number rather than a fortnight-old one.

Choose Ideogram 4.5 if your deliverable is print-resolution or source-resolution, and if the specific thing breaking your workflow is that the tenth edit has ruined the first. That is a real problem and 4.5 is the only shipped product addressing it directly. Go in knowing that you are paying roughly 3.7 times the Grok rate for a property no one has independently measured, and that the property it is not selling — raw editing preference — currently scores 43 Elo in xAI's favour.

The tiebreaker is the same test in both cases, and it takes an afternoon: one image, ten edits, diff the untouched regions between turn one and turn ten. Run it on the cheap model first. If it holds, you have saved yourself a specialist contract. If it falls apart, you now know exactly which images Ideogram 4.5 is for — and you have the before-and-after to justify the $0.22.