Ideogram 4.5 vs Flux 3: A Shipped Editor Against an Unshipped Modality

Ideogram 4.5 and Flux 3 are the two most recent frontier image announcements that share a thesis, and only one of them is a thing you can call. Ideogram 4.5 went live on September 30, 2026 as a precise-editing model built around a specific claim: repeated edits stop degrading the image. Flux 3, Black Forest Labs' unified multimodal foundation model, was announced on July 23, 2026 and promised the same property for its image tier — non-destructive editing that preserves texture, grain and lighting across turns. Ten weeks later, Flux 3 Image still has no public endpoint, no model identifier and no rate card. What BFL has actually shipped from that family is video and a robotics action model.

So the honest form of this comparison is not "which is better." It is "one is a priced product with an independent score, and the other is a promise with a deadline that has passed." Everything below separates the two.

What you can call today

• Ideogram 4.5 — live since September 30, 2026 on Ideogram's own app and API, plus partner platforms. Two operations, generate and precise edit, at four rendering speeds. Independent rows exist on Artificial Analysis's text-to-image board (34th of 167, 1,014 Elo) and its editing board (23rd, 1,064 Elo).

• Flux 3 Image — no public endpoint. Black Forest Labs' own documentation recommends FLUX.2 for text-to-image and image editing and says so in plain language: "FLUX.2 is the recommended model for text-to-image generation," and for editing, "for new projects, we recommend FLUX.2 with multi-reference support and up to 4MP output." The company's Flux 3 material is video and action.

• Flux 3 Video — general availability since August 4, 2026, extended on September 10 with 2K and 4K output modes. Billed per second of output: $0.17/s at HD, rising to $0.80/s for UHD text-to-video and $0.95/s for UHD video-to-video.

• Flux 3 Action — a 7B open-weights world-action model, shipped, with documentation covering fine-tuning, a SO-101 robot arm recipe, and examples for drone control and game playing.

• The callable BFL image line — FLUX.2 in its five sizes: [pro], [max], [flex], [dev] and [klein]. This is the previous generation and it is what BFL itself points production image work at.

Three of the five modalities Flux 3 was announced with have shipped. The one that competes with Ideogram 4.5 has not, and BFL is routing its own customers to the generation before it in the meantime. That is the single most consequential fact in the matchup.

The claim both launches are actually making

Both companies are selling the same technical property, and it is not image quality. It is durability.

Image editors have a characteristic failure: ask for one small change and the model re-renders the frame. Most of the output survives. The rest does not, and the difference compounds with every pass — faces shift, grain changes, text softens, lighting drifts. Anyone who has done ten-turn work in a chat interface has watched it happen.

Ideogram 4.5's answer is engineering at the edit boundary. The model, its page says, "preserves the edges of an edited crop, so you can stitch it seamlessly back into the original," demonstrated on a 4,016 × 6,016 source — 24.2 megapixels — edited without being downscaled first. The pitch is fidelity preserved, not taste improved.

Flux 3's answer, as roadmapped in July, was the same shape: "non-destructive editing — preserving textures, grain, lighting, and background so multi-turn edits do not degrade." It appears in a feature list for the image tier. The image tier has not arrived, so the claim has never been tested by anyone outside Black Forest Labs.

That is worth pausing on. Two labs independently concluded that drift is the hard problem. One shipped a mitigation and put a price on it. The other described the mitigation and has not yet put the model in front of a customer. If you are choosing today, you are choosing between a working answer and an anticipated one.

What the scoreboards do and do not say

Artificial Analysis runs blind human-preference arenas for both text-to-image and editing, and Ideogram 4.5 has entries on each. The figures, taken from the boards in the first days of October 2026:

• Text-to-image — 34th of 167 models at 1,014 ± 11 Elo, 2,247 votes, $100.00 per 1,000 images. The leader is GPT Image 2.5 Sunburst (max) at 1,197. Ideogram's own 4.0 from June scores 1,011, so the generational gain on this board is roughly three Elo.

• Image editing — 23rd at 1,064 ± 11 Elo, 2,382 votes, $220.00 per 1,000 images on the board's conversion. The leaders are GPT Image 2.5 Sunburst at 1,182 and GPT Image 2.5 Flare at 1,162.

• Flux 3 — absent from both. No row exists because no endpoint exists to test. The Flux entries on the text-to-image board are all FLUX.2: [flex] at 1,028, [max] at 1,021, [pro] at 1,004, [dev] at 1,000, and on the editing board [klein] 9B at 1,013.

Two caveats belong with those numbers. Ideogram 4.5's rows are new and thin — a few thousand votes against the 12,000 to 18,000 the leaders carry — which is why its interval is ±11 rather than ±8. And neither board asks voters about drift. The property Ideogram is selling is invisible to a single-shot preference test, so a 23rd-place editing rank measures something adjacent to the claim rather than the claim itself.

The price side, which is where problems start

• Ideogram 4.5 — $0.03 per image at Low, $0.06 at Medium, $0.22 at High on the reseller rate cards published at launch; the same $0.22 ceiling appears in Artificial Analysis's $220.00-per-1,000 figure for the 2K High configuration. Cost tracks the rendering-speed setting.

• Flux 3 Image — nothing published, because nothing exists. BFL's only Flux 3 pricing is per-second video: $0.17/s HD text-to-video, $0.40/s at 2K, $0.80/s at 4K, with video-to-video at $0.65/s and $0.95/s in the same bands.

• The callable alternative — FLUX.2 image rates on the same leaderboard run from $12.00 per 1,000 for [dev] through $30.00 for [pro] and $70.00 for [max]. That is the line you would compare Ideogram 4.5 against if you are choosing between two things that exist, and on raw Elo-per-dollar Ideogram 4.5 does not win: 1,014 Elo at $100 per 1,000 against FLUX.2 [flex]'s 1,028 at $60.

Which is the honest framing of the whole matchup. Against the model BFL actually serves, Ideogram 4.5 is not a better generator and not a cheaper one. It is a different product that does one thing FLUX.2 does not advertise as a headline, at a price that only makes sense if that one thing is your bottleneck.

Reading the ten-week gap

The July 23 announcement said image access would follow "after an early access phase," and the company indicated that early access would open "in the following weeks." It is now October 1. In that window BFL shipped Flux 3 Video to general availability, extended it to 2K and 4K, added an MCP server for video generation, published agent skills, capped video continuation at 15 seconds, added auto top-up for credits, and brought a separate open-weights action model to release.

That is a company shipping steadily — just not the image half. The plausible read is prioritisation rather than trouble: a unified multimodal model is a sequencing problem, video was the differentiator worth leading with, and the image tier is the modality where BFL already has an incumbent product (FLUX.2) that would be cannibalised. The natural reading of "we recommend FLUX.2 for text-to-image" in the current docs is that the image tier is waiting its turn.

What that means for a buyer is simple. If you need an editor with a drift claim today, Ideogram 4.5 is the only shipped answer from either lab, and the price is $0.22 per image at the setting that claim is demonstrated at. If you need the best raw generator, neither of these is the right comparison — the arena leaders are OpenAI's 2.5 pair and, at the value end, FLUX.2 [flex] at $60 per 1,000.

Evaluating this without betting on it

The decision most teams actually face is not "which of these two," because there is only one. It is "is Ideogram 4.5 worth adding to a stack that already generates images, for the editing case specifically."

The test is cheap and the only one that speaks to the claim: take ten images your current pipeline handles badly, run the same multi-turn edit sequence on each with Ideogram 4.5, and diff turn one against turn ten. Measure the divergence. Price it per completed edit rather than per generated image, because at High on a large source you are paying $0.22 a pass. If the divergence is small enough to skip a manual repair step, the model has paid for itself on that workflow alone; if it is not, the arena score was telling you the truth about a model that is expensive for what it does.

Running that experiment alongside the model you already use is the case for a routing layer rather than another direct integration — one endpoint and one key covering the incumbents and the new entrant, per-call pricing with no markup added on top, and automatic failover so a specialist trial cannot take down a production path. It is a lower-stakes way to answer a question that no leaderboard is going to answer for you.

Where this lands

Ideogram 4.5 is a real release with a real price and a claim nobody has yet been able to falsify. Flux 3 Image is a real roadmap item with no release, and the ten weeks since its announcement have been filled with shipped work on every modality except that one. Comparing them on image quality is comparing a product to a plan.

The thing to watch is not the next Flux 3 release note — it is whether BFL ever attaches an endpoint, an identifier and a rate to the image tier. If it does, this becomes a genuine head-to-head between two laboratories that diagnosed the same problem. If it does not, Ideogram 4.5 stands alone as the only shipped answer to multi-turn edit drift, and its mid-table arena rank will keep looking like the wrong number to judge it by.