No image model is best at every job, and the vendors say so about their own. Seven of them publish what their model is for, where it falls short and what it costs. This page lines those statements up by job, from the vendors’ own pages read on 18 September 2026. Nothing was generated or tested here: it is a map of claims, not a ranking.
- Text in images: Google and Black Forest Labs claim it most strongly; OpenAI, ByteDance and Meta publish caveats.
- Editing a photo: every vendor here edits an uploaded image; Midjourney’s edit model is still in testing.
- Consistent characters: Google publishes limits, Meta a method, and OpenAI a warning.
- Cheapest bulk: Meta’s Muse Image has the lowest flat per-image price. OpenAI prices by token; Midjourney sells subscriptions.
- Run it yourself: only Black Forest Labs offers downloadable weights here, and only its smallest model is free for commercial use.
- Negative prompts: only Midjourney documents one.
Every price here was correct when this site last checked it, between 18 Sep and 22 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: Meta pricing · Black Forest Labs pricing · BytePlus pricing · Google pricing · xAI pricing · OpenAI pricing.
Nine models, side by side
Each cell is the vendor’s wording or a plain summary of its documentation. Prices are the lowest per-image API price published, a snapshot from 18 Sep 2026; bigger or higher-quality images cost more.
| Model | What the vendor says it is for | Negative prompt | Text in the image | Editing and references | Lowest API price per image |
|---|---|---|---|---|---|
| ChatGPT Images 2.5 (API: gpt-image-2.5-sunburst, -flare) | Sunburst “for workflows where editing precision matters most”, Flare “for fast, high-quality everyday image generation” | No field; exclusions go in the prompt | “can still struggle with precise text placement and clarity” | Edits, reference images and masks | Priced per token; OpenAI’s calculator estimates $0.00588 at low quality |
| Nano Banana 2 (gemini-3.1-flash-image) | “generalist workhorse model for all tasks” | No field; “semantic negative prompts” | “reliable text rendering” | Up to 14 reference images | $0.067 at 1K |
| Nano Banana 2 Lite | “Our fastest and cheapest Gemini image model” | As above | No separate claim | “Not optimized for multiple reference inputs or multi-turn sequential editing.” | $0.0336 at 1K |
| Nano Banana Pro | “The premium choice for the most complex visual tasks” | As above | “Use Gemini 3 Pro Image for professional asset production.” | Up to 14 reference images | $0.134 at 1K to 2K |
| Midjourney V8.2 | “an update focused on aesthetics, image quality, and Personalization” | Yes: --no | No claim on the pages read | Edit model in testing, “up to 4 image references at once” | None: subscriptions from $10 to $120 a month |
| FLUX.2 [klein], [pro], [flex], [max] | [klein] “real-time, high-volume generation”; [max] “maximum quality and grounding search” | No: “FLUX.2 does not support negative prompts.” | [flex] “Specialized for typography.” | Up to 8 references by API on [pro], [flex] and [max]; up to 4 on [klein] | From $0.014 ([klein] 4B); from $0.03 ([pro]) |
| Seedream 5.0 Pro, 5.0 Lite and 5.0 Flash | 5.0 Pro: “Complex information visualization” | Not in the guides read | Improved, with “still room to improve in finer-grained text rendering” | Point and lasso selection, layer separation; up to six references on 5.0 | $0.018 (Flash); $0.035 (Lite); $0.045 (Pro, up to 2.61 million pixels) |
| Grok Imagine (grok-imagine-image-2.0) | “Generate images from text prompts with Grok Imagine models.” | Not documented | No claim on the pages read | Chained edits; “up to five source images for editing” | $0.04 (2.0, 1K, low); $0.02 (grok-imagine-image, 1K) |
| Muse Image (Meta) | “Agentic image generation, grounded by search and priced for production volumes” | Not documented on the pages read | “Exact text rendering varies run to run.” | Edits, multi-image composition, “anchored composition” | $0.01 |
OpenAI, Image generation and Image prompting · Google, Nano Banana image generation and Gemini Developer API pricing (read in a browser; they refuse automated fetches) · Midjourney, Version, No, Edit Model for V8 and Comparing Midjourney Plans · Black Forest Labs, FLUX.2 overview, FLUX.2 prompting guide and API Pricing · ByteDance, Introducing Seedream 5.0 Pro, Dreamina How to use Seedream 5.0 and BytePlus ModelArk Pricing (read in a browser) · xAI, Image Generation, Image Editing, Imagine Image Quality Retirement and Pricing · Meta, Muse Image and Build with Muse Image (read in a browser). All read 18 Sep 2026; the Seedream 5.0 Flash price read in a browser 22 Sep 2026. The ChatGPT Images 2.5 name for the gpt-image-2.5 models is from this site’s cheat sheet, checked 16 Sep 2026.
Only Midjourney documents a negative prompt, and even --no is a weight: Midjourney says it “is the same as weighing part of a multi-prompt to” minus 0.5. Black Forest Labs says “FLUX.2 does not support negative prompts.” Writing exclusions as descriptions instead is covered on negative prompts.Midjourney, No, read at source 18 Sep 2026: “The --no parameter is a tool you can use to tell Midjourney what you don't want in your image.” · Black Forest Labs, FLUX.2 prompting guide, read at source 18 Sep 2026: “FLUX.2 does not support negative prompts.”
Several vendors publish what their model is not for. Read that first: they had no reason to write it.
Text in images: posters, labels, menus
Google makes the plainest claim, “Gemini excels at rendering text”, and adds a method: “Gemini works best if you first generate the text and then ask for an image with the text.”Google, Nano Banana image generation, read in a browser 18 Sep 2026: “Gemini excels at rendering text.”
Black Forest Labs sells one FLUX.2 tier for this job, describing [flex] as “Specialized for typography. Best for text rendering and preserving small details.”Black Forest Labs, FLUX.2 overview, read at source 18 Sep 2026: “Specialized for typography. Best for text rendering and preserving small details.”
ByteDance pitches Seedream 5.0 Pro at dense layouts and, in the same post, says there is “still room to improve in finer-grained text rendering and pixel-level editing consistency”.ByteDance Seed, Introducing Seedream 5.0 Pro, read at source 18 Sep 2026: “there is still room to improve in finer-grained text rendering and pixel-level editing consistency”.
OpenAI and Meta both lead with the limit. OpenAI: “Although significantly improved, the model can still struggle with precise text placement and clarity.” Meta: “Exact text rendering varies run to run.”OpenAI, Image generation, read at source 18 Sep 2026: “Although significantly improved, the model can still struggle with precise text placement and clarity.” · Meta, Build with Muse Image, 25 Aug 2026, read in a browser 18 Sep 2026: “Exact text rendering varies run to run.”
Midjourney and xAI make no text claim on the pages read. Prompting for words is on text inside images.
For words in a picture, start where a vendor makes a claim, and still proofread every letter.
Editing a photo you already have
Every vendor here accepts an input image. They differ in how you point at the part to change.
- OpenAI takes a mask, and steers editing to Sunburst, its model for “workflows where editing precision matters most”.OpenAI, Image generation, read at source 18 Sep 2026: “Choose Sunburst for workflows where editing precision matters most, and Flare for fast, high-quality everyday image generation.”
- Google does it in conversation: “Provide an image and describe your change.”Google, Nano Banana image generation, read in a browser 18 Sep 2026: “Provide an image and describe your change.”
- Seedream 5.0 Pro gives you tools: “point selection, lasso selection, sketch rendering, color and material replacement, layer separation, and multi-image fusion”.ByteDance Seed, Introducing Seedream 5.0 Pro, read at source 18 Sep 2026: “point selection, lasso selection, sketch rendering, color and material replacement, layer separation, and multi-image fusion”.
- Grok Imagine works in steps: “Chain multiple edits together by using each output as the input for the next.”xAI, Image Editing, read at source 18 Sep 2026: “Chain multiple edits together by using each output as the input for the next.”
- Muse Image “is designed to change exactly what you specify and leave everything else untouched.”Meta, Muse Image, read in a browser 18 Sep 2026: “Muse Image is designed to change exactly what you specify and leave everything else untouched.”
- Midjourney opened its V8.2 edit model, with inpainting and outpainting, for testing on 27 Aug 2026.Midjourney, Edit Model for V8, 27 Aug 2026, read at source 18 Sep 2026: “start letting everyone test our first V8.2 image edit model”.
If you must say exactly which pixels change, pick a model that takes a mask or a selection. If describing the change is enough, any of them will try.
Keeping one character consistent
Google publishes numbers: gemini-3.1-flash-image “supports character resemblance of up to 4 characters and the fidelity of up to 10 objects in a single workflow”, and the Gemini 3 image models take up to 14 reference images. Its Lite model is the one to avoid for this job.Google, Nano Banana image generation, read in a browser 18 Sep 2026: “gemini-3.1-flash-image supports character resemblance of up to 4 characters and the fidelity of up to 10 objects in a single workflow.” and, of Nano Banana 2 Lite, “Not optimized for multiple reference inputs or multi-turn sequential editing.”
Meta documents a method, “anchored composition”: “Anchor a whole series of generations on a small set of reference images so the character, style, and setting stay consistent from one image to the next.”Meta, Build with Muse Image, 25 Aug 2026, read in a browser 18 Sep 2026: “Anchor a whole series of generations on a small set of reference images so the character, style, and setting stay consistent from one image to the next.”
Black Forest Labs allows up to 8 references by API on [pro], [flex] and [max], and says its open FLUX.2 [dev] handles “character, object and style reference without additional training”. ByteDance lets Seedream 5.0 users “upload up to six reference images”.Black Forest Labs, FLUX.2 overview, read at source 18 Sep 2026: multi-reference “Up to 8 (API), 10 (playground)” · FLUX.2 [dev] model card, read at source 18 Sep 2026 · Dreamina, How to use Seedream 5.0, read at source 18 Sep 2026: “Users can upload up to six reference images for complex style fusion”.
OpenAI says its model “may occasionally struggle to maintain visual consistency for recurring characters or brand elements across multiple generations.”OpenAI, Image generation, read at source 18 Sep 2026, under Limitations. Techniques that work across models are on character consistency.
Consistency comes from references, not luck. Choose by how many a model takes and whether its vendor says it is built to hold them.
Cheapest for bulk work
On flat per-image list prices, Meta’s Muse Image is lowest at $0.01, then FLUX.2 [klein] 4B from $0.014, ByteDance’s Seedream 5.0 Flash at $0.018 and the older grok-imagine-image at $0.02. Google’s Nano Banana 2 costs $0.067 at 1K, and Nano Banana Pro $0.134.Meta, Muse Image, read in a browser 18 Sep 2026: “priced for production volumes at $0.01/image” · Black Forest Labs, API Pricing, read at source 18 Sep 2026: FLUX.2 [klein] 4B “from $0.014”, FLUX.2 [pro] “from $0.03” · xAI, Pricing, read at source 18 Sep 2026: grok-imagine-image-2.0 “$0.04 / img” at 1K low, grok-imagine-image “$0.02 / img” at 1K · ByteDance, BytePlus ModelArk Pricing, read in a browser 18 Sep 2026: seedream-5-0-lite “0.035”, Seedream 5.0 Pro “≤ 2.61 million pixels (1.5K or lower): 0.045” · Google, Gemini Developer API pricing, read in a browser 18 Sep 2026: “equivalent to $0.067 per image” (Nano Banana 2, 1K), “equivalent to $0.0336 per image” (2 Lite, 1K), “equivalent to $0.134 per image” (Pro, 1K to 2K) · OpenAI, Image generation, read at source 18 Sep 2026: calculator “$0.00588”, “Per image at 30 USD per million image output tokens. Excludes text and image input tokens and streaming partial images.”
Seedream 5.0 Flash is now the cheapest Seedream. It appears on BytePlus’s price list as dola-seedream-5-0-flash-260915 at $0.018 per output image, with input images free, about half the $0.035 of 5.0 Lite. This page found no ByteDance announcement of the model, only the price-list row, so what it is for is not stated.ByteDance, BytePlus ModelArk Pricing, page last updated 22 Sep 2026, read in a browser 22 Sep 2026: dola-seedream-5-0-flash-260915, input image “Free”, output image “0.018” (USD per image).
Two vendors do not fit a per-image column. OpenAI prices GPT Image 2.5 at $30 per million image output tokens; its calculator, on its defaults of low quality and 1024 by 1024, estimates $0.00588 per image before input tokens — under Meta’s cent, but an estimate at the lowest setting, not a list price. Midjourney sells no images: “All Midjourney plans are subscriptions”, from $10 to $120 a month, and the one API page found is a July 2025 post about starting to “investigate opening up an Enterprise API”.OpenAI, Image generation, read at source 18 Sep 2026: “$30 per million image output tokens” · Midjourney, Comparing Midjourney Plans, read at source 18 Sep 2026: “All Midjourney plans are subscriptions” · Midjourney, Enterprise API Survey, 16 Jul 2025, read at source 18 Sep 2026.
Two older models have end dates. Google’s original Nano Banana, gemini-2.5-flash-image, has two, depending on where you call it: on the Gemini Developer API it “is deprecated and will be shut down on October 2, 2026”, while on Google Cloud’s Gemini Enterprise Agent Platform it is now “Deprecated and scheduled for retirement on March 15, 2027 (extended from October 2, 2026)”. And xAI retires grok-imagine-image-quality on 2 November 2026, after which that name is served by 2.0 at low quality rather than failing.xAI, migration guide, read at source 22 Sep 2026: “After November 2, 2026, requests to grok-imagine-image-quality are served by grok-imagine-image-2.0 with quality set to low. The slug continues to resolve, so you do not need to change your code to avoid breakage.” · Google, Gemini Developer API pricing, read in a browser 18 Sep 2026 and again 22 Sep 2026: “is deprecated and will be shut down on October 2, 2026” · Google Cloud, Gemini Enterprise Agent Platform release notes, 14 Sep 2026 entry, read in a browser 22 Sep 2026: “Gemini 2.5 Flash Image (gemini-2.5-flash-image): Deprecated and scheduled for retirement on March 15, 2027 (extended from October 2, 2026).” · xAI, Imagine Image Quality Retirement, read at source 18 Sep 2026: “Effective November 2, 2026, the grok-imagine-image-quality model slug is retired from the xAI API.”
The sticker is not the bill. A cheap model that needs four attempts costs more than a dearer one that works first time; the arithmetic is on what generation really costs, and an earlier three-vendor comparison is on Meta’s developer stack.
Compare prices at the size and quality you will ship, then divide by your hit rate. The lowest number on a pricing page is usually the smallest setting.
Open weights to run yourself
Among the models in the table, FLUX is the one with weights you can download; the other six vendors’ pages read here offer theirs through an app or an API. Licences differ within the family. FLUX.2 [klein] 4B has “Open weights available for commercial use under the Apache 2.0 license” and, by its model card, runs on consumer hardware “with as little as 13GB VRAM”.Black Forest Labs, FLUX.2 [klein] 4B model card, read at source 18 Sep 2026: “Open weights available for commercial use under the Apache 2.0 license.”
The larger open models are not free for business use: “[klein] 9B is available under the FLUX Non-Commercial License”, and of FLUX.2 [dev], “This model falls under the FLUX Non-Commercial License.” The [dev] card adds that “Generated outputs can be used for personal, scientific, and commercial purposes”, so the restriction is on running the model, not on the pictures. BFL sells commercial self-hosting licences.Black Forest Labs, FLUX.2 overview and FLUX.2 [dev] model card, read at source 18 Sep 2026: “This model falls under the FLUX Non-Commercial License.” · Pricing, read at source 18 Sep 2026: “Deploy and run FLUX models on your own infrastructure.” The wider trade-off is on local or cloud.
“Open” is a licence, not a price. Check which FLUX model you downloaded before client work.
Free or paid? What each tool’s free plan allows, and who owns the output: free vs paid AI image, video and music tools.
What this page could not verify
- Quality. Every capability here is the vendor’s claim; nothing was generated or compared.
- OpenAI’s cost above low quality. The calculator would not change setting in the browser used, so only its default is reported.
- Whether Muse Image’s $0.01 covers every size. Meta gives one figure; its guide says output goes “up to 1600px”.Meta, Muse Image and Build with Muse Image, read in a browser 18 Sep 2026: “Output is high-quality up to 1600px.”
- Negative prompts on Grok Imagine, Seedream and Muse Image. Not mentioned either way on the pages read.
- A Midjourney API. No current API price was found; the one page found is the July 2025 survey.
- Tomorrow’s prices. Every price is a snapshot from 18 Sep 2026; re-read before you budget.
OpenAI, Image generation, Image prompting and Pricing · Google, Nano Banana image generation and Gemini Developer API pricing · Midjourney, Version, No, Comparing Midjourney Plans, Edit Model for V8 and Enterprise API Survey · Black Forest Labs, FLUX.2 overview, FLUX.2 prompting guide, API Pricing, Pricing, and the FLUX.2 [klein] 4B and FLUX.2 [dev] model cards · ByteDance, Introducing Seedream 5.0 Pro, How to use Seedream 5.0 and BytePlus ModelArk Pricing · xAI, Image Generation, Image Editing, Imagine Image Quality Retirement and Pricing · Meta, Muse Image, Build with Muse Image and AI developer site. All read on 18 September 2026; Google, BytePlus and Meta pages in a browser because they refuse automated fetches.
Prompting rules for each model are on the model cheat sheets.