Ideogram 4.5 vs Meta Muse Image: A Precision Model and a Penny Model, Priced From Opposite Ends

Ideogram 4.5 arrived on 30 September 2026 with a single-sentence pitch — "the most precise edit model" — and a rate card that starts at 0.8 US cents an image and ends at 22 cents. Meta Muse Image has been live since 7 July 2026 at a flat one cent, no tiers, no ladder. Put the two next to each other and the price gap between Ideogram 4.5's cheapest tier and Muse Image is two-tenths of a cent, while the gap between the two models at their most expensive is twenty-two to one. That ladder, not the model cards, is the first thing worth understanding: one of these is sold as a specialist priced by how hard you are working it, and the other is sold as an agent that costs the same whether you ask for a thumbnail or a poster.

The awkward part, and the reason this comparison cannot be written as a simple quality verdict, is that Meta Muse Image is currently measured higher than Ideogram 4.5 on the independent board that matters most for an edit model. On Artificial Analysis's image-editing leaderboard Muse Image sits sixth at an Elo of 1,118 from 12,272 votes; Ideogram 4.5 (High) is twenty-third at 1,064 from 2,382. Ideogram's own page does not dispute this — it makes no numeric claim at all against Muse. It makes a video.

What Ideogram 4.5 is, precisely

Ideogram 4.5 is a precise-edit model, which in this generation means it is judged on what survives an edit rather than on what a single generation looks like. Ideogram's stated problem is drift: "with each edit, image models can introduce pixel shifts, color changes, and texture artifacts." The claimed fix is that 4.5 "reduces this drift, preserving details across multi-turn edits" — turn one is clean, turn six is still clean.

The second claim is resolution. Ideogram demonstrates a colourway edit performed on a 4,016 × 6,016 pixel source, 24.2 megapixels, with the model preserving the edges of the edited crop so it can be stitched back into the original without downsizing the rest of the image. That is a different engineering problem from generating 4K from scratch, and it is the one that matters if you are retouching a real campaign asset rather than making a new one.

Around those two claims Ideogram lists ten edit categories it says 4.5 handles with "exceptional precision": colour and lighting, text modification, product photography, interior design, architecture, restoring old photographs, sketch to image, style reference, reframe, and depth to image. The text-modification examples are the ones worth noticing for this blog's usual audience — brand systems translated into Japanese, French and English while keeping the layout, which is the job typography-led pipelines have been paying a premium for.

Meta Muse Image is the other archetype. It is Meta's image model, launched to consumers on 7 July 2026 across Meta's own surfaces and to developers afterward on Meta's model API at a flat $0.01 per generated image, and it is built as an agentic image model rather than a precision editor: it takes up to ten reference images in a single request, reasons about the request before rendering, and is deliberately undifferentiated on price. Output tops out well below Ideogram 4.5's demonstrated working resolution.

The rate card, and what a ladder costs you

The Ideogram pricing page is unusually explicit, and it is worth reading as a shape rather than as five numbers. Ideogram 4.5 is billed per generated image, and the price is a function of two things: rendering speed and output size. Speed has four steps and the price moves 12.5× across them; resolution, from 1K to 2K, changes nothing at all.

• Price at Very Low — $0.008 per image, at both 1K and 2K

• Price at Low — $0.03 per image, at both 1K and 2K

• Price at Medium — $0.06 per image, at both 1K and 2K

• Price at High — $0.10 per image, at both 1K and 2K

• Price when the request includes a source image — High rises to $0.22, and the Precise Edit endpoint prices the same: $0.008 Very Low, $0.03 Low, $0.06 Medium, $0.22 High

• Meta Muse Image — one price, $0.01 per generated image, regardless of speed or size

Two practical consequences fall out of that shape. First, image-to-image and precise editing are not a separate product with a separate price at the low end — a Low edit costs the same $0.03 as a Low generation, and only the High tier carries a distinct $0.22. Second, the ladder means your effective cost is a function of your own tolerance, not of the model's quality tier: a pipeline that accepts a Very Low render pays eighty times less than one that wants High with a source image, for the same model.

Against Meta that is a straightforward trade. If you generate a flat mix of small images at low settings, Ideogram 4.5 is cheaper than a penny and Muse Image is a penny. If your work is precision work, Ideogram 4.5 costs up to twenty-two times as much, and you are buying the resolution and the drift claim rather than a measured quality advantage.

Where the independent evidence actually points

This is the section Ideogram's launch page does not have, and it is short, because Ideogram 4.5 is one day old and Artificial Analysis has just started collecting on it.

On the editing board, Muse Image is sixth at Elo 1,118 with 12,272 votes collected since July. Ideogram 4.5 (High) is twenty-third at 1,064 with 2,382 votes. Those error bars are wide on both sides — 1,110 to 1,126 for Muse, 1,053 to 1,075 for Ideogram — but they do not overlap, so the current ordering is real if provisional. On the text-to-image board the same pattern holds with more separation: Muse Image seventh at 1,114, Ideogram 4.5 (High) thirty-fourth at 1,014.

There is one genuinely interesting detail in those two numbers. Artificial Analysis lists Ideogram 4.5 (High) at $100 per thousand images on the text-to-image board and $220 per thousand on the editing board, which is exactly Ideogram's own $0.10 and $0.22 tiers. The measurement board and the rate card agree, which tells you the person running that board is calling the real API at the real price rather than benchmarking a demo build.

For context on whether 4.5 is an upgrade worth paying for: Ideogram 4.0 (Quality) sits at 1,011 on text-to-image from 12,176 votes, and 4.0 does not appear on the editing board at all. So 4.5 is Ideogram's first entry in AA's editing arena, and its text-to-image number is a three-point improvement on a predecessor with five times the votes. Neither of those is a large enough delta to settle anything, and it is genuinely too early to say more.

What Ideogram's own comparison proves, and what it does not

Ideogram's launch page runs an explicit comparison against three other models: GPT Image 2.5 Sunburst, Nano Banana Pro and Nano Banana 2. The claim is strong — "GPT Image and Nano Banana outputs become unusable within a few edits, while Ideogram 4.5 stays clean edit after edit."

The evidence offered for that claim is a video montage of the same edits applied to each model, chosen by Ideogram, presented without a metric, a prompt set, a vote count or a judge. That is not a criticism of the model; it is a statement about what the artifact supports. It is the single most decision-relevant claim Ideogram makes — multi-turn stability is the whole product — and it is currently the least measurable, because no independent arena tests multi-turn editing drift as its own axis. The AA editing board scores single-edit quality, which is why Muse Image can rank sixth on it while Ideogram's comparison shows it degrading across turns. Those two facts are not in conflict. They measure different things, and only one of them has votes.

The resolution demonstration is different: it is a specific, checkable artifact, a 24.2-megapixel source and an edit that preserves crop edges. That claim is corroborated by the pricing structure — the $0.22 High-with-source tier exists precisely because in-place editing at that size is the expensive operation — and it is the claim competitors' comparator pages, most of which are written against Ideogram 4.0 and not 4.5, have not yet caught up to.

Getting either of them

Neither model is on OrcaRouter today. Ideogram sells 4.5 through its own app and its own API — the vendored endpoints are /v2/image/generate/ideogram-4-5 and /v2/image/precise-edit/ideogram-4-5 — plus launch partners we are not naming here. Meta Muse Image is on Meta's own model API and, again, third-party platforms.

What we do route, and what makes that worth saying in an edit-model comparison, is the other side of the same job. OrcaRouter carries 200+ models behind one OpenAI-compatible endpoint with provider list price passed through at 0% markup, and that includes the image models people actually compose with: google/gemini-3.1-flash-image-preview at $0.151 a call, google/gemini-3-pro-image-preview at $0.24, grok/grok-imagine-image at $0.20, and the GPT-Image line. Automatic failover matters more in image work than in text, because a render that fails after ninety seconds is ninety seconds you cannot get back, and the routing DSL lets you pin a specific model per request instead of maintaining an SDK per vendor. If you are deciding between the two models in this article, the useful thing an aggregator gives you is not access to either of them — it is a stable, priced baseline to measure both against, and a place to send the jobs that do not need a precision edit at all.

Which one to reach for

If your workload is a single generation at a moderate size, repeated at volume, Muse Image is the better buy on both evidence and price. It is measured ahead on both boards, it takes ten reference images in one request, and at a flat cent it is cheaper than every Ideogram 4.5 tier except Very Low. You give up working resolution above roughly 1600 pixels and you give up the tiering that lets you pay less for a draft.

If your workload is subtractive rather than generative — take an existing 24-megapixel asset, change one element, keep the other twenty-four million pixels exactly as they were — Ideogram 4.5 is the only one of the two that claims to do it, and the $0.22 Precise Edit High price is the cost of that claim. What you are accepting in exchange is that the claim's central piece, multi-turn stability, has no independent measurement at all, that the model's overall editing quality currently measures twenty-three points of Elo below the penny model, and that its published evidence for the headline comparison is a video rather than a table.

The thing to watch is not Ideogram's leaderboard position — with 2,382 votes on a day-old model it will move, and probably up, since AA's editing board is the one where the model was built to compete. The thing to watch is whether anyone builds a multi-turn drift benchmark. Ideogram 4.5's whole thesis, and the $0.22 tier that prices it, depends on a property nobody outside the lab has yet measured.

OrcaRouter runs 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, which in this matchup is the difference between one key and two vendor accounts.