# GPT Image 2.5, Nano Banana, Seedream, Flux 3 and Grok join eroq

> Seven new image models on eroq: GPT Image 2.5 Sunburst and Flare, Nano Banana Pro and 2.1, Seedream 5.0 Pro, Flux 3 and Grok Imagine 2.0, and what each is for.

Published 2026-10-08 · eroq.ai — canonical: https://eroq.ai/blog/new-image-models-gpt-image-nano-banana-seedream-flux-grok


Seven image models from five makers joined eroq in October 2026. OpenAI brings two, [GPT Image 2.5 Sunburst](/models/gpt-image-2-5-sunburst) and [GPT Image 2.5 Flare](/models/gpt-image-2-5-flare). Google brings [Nano Banana Pro](/models/nano-banana-pro) and [Nano Banana 2.1](/models/nano-banana-2-1). ByteDance adds [Seedream 5.0 Pro](/models/seedream-5-0-pro), Black Forest Labs adds [Flux 3 Image](/models/flux-3-image) and xAI adds [Grok Imagine 2.0](/models/grok-imagine-2-0). They sit in the same image studio as eroq's own engines, which takes the lineup to 13 image models, and every account can use them today, pay-as-you-go included.

The one-line version: Sunburst when the frame has words in it, Flare when it needs drama, Nano Banana Pro when the final has to be right, Nano Banana 2.1 when you want a good draft now, Seedream for texture, Flux 3 for natural light, and Grok for a look you would not have thought to ask for.

## The seven at a glance

| Model | Maker | Price | Typical render | Reach for it when |
|---|---|---|---|---|
| GPT Image 2.5 Sunburst | OpenAI | 31 credits / image | about 15 s | the frame carries words or a strict layout |
| GPT Image 2.5 Flare | OpenAI | 31 credits / image | about 15 s | you want key art, a poster, a cinematic still |
| Nano Banana Pro | Google | 124 credits / image | about 25 s | the final has to hold every detail |
| Nano Banana 2.1 | Google | 31 credits / image | about 8 s | you want a good image now, in plain English |
| Seedream 5.0 Pro | ByteDance | 36 credits / image | about 20 s | texture and dense detail are the point |
| Flux 3 Image | Black Forest Labs | 32 credits / image | about 20 s | natural light, photo realism, typography |
| Grok Imagine 2.0 | xAI | 32 credits / image | about 15 s | you are exploring many directions |

Times are the studio's planning estimates, not a promise: a busy hour adds queue time, and the result panel shows how long each take really took.

## What all seven have in common

The ground rules are identical across the batch:

- **Prompt in, picture out.** All seven are text-to-image on eroq: no reference pictures, no Edit mode.
- **Five formats at one price.** 1:1, 3:4, 4:3, 9:16 and 16:9, each rendered natively.
- **One to four takes per run.** Each take is billed and refunded on its own: if one take of four fails, you pay for three.
- **No plan gate.** A plain credit pack runs every one of them.
- **Moderated upstream.** Each provider applies its own content rules. A prompt it refuses comes back as `content_blocked` and costs nothing.
- **Privacy, stated plainly.** Outside providers run them; they may retain your prompt but never use it for training. With Private mode on, the result is never stored.

## Pick one in thirty seconds

- **The image has to say something** (a headline, a label, a shop sign): Sunburst. Flux 3 when the words sit on a real surface in a photograph; Nano Banana Pro when they sit inside a dense, detailed scene.
- **It is a poster, a cover or a still from a film that does not exist:** Flare.
- **It is a campaign hero that will be seen large:** Nano Banana Pro, once you have found the idea on something cheaper.
- **You are still working out what the picture is:** Nano Banana 2.1 for plain-English iterations, Grok Imagine 2.0 for characterful variations.
- **The subject is a material or a place** (a canyon wall, a cathedral nave, a hand-knit sweater): Seedream 5.0 Pro.
- **It should look like a photograph someone took, not a staged render:** Flux 3 Image.
- **It must match a real face, product or outfit from your own pictures:** none of the seven. Read the next section.

The [buyer guide to every image model](/blog/best-ai-image-models-2026) runs this exercise across the whole lineup.

## What they do not do, and what does

None of the seven reads your pictures. When identity matters (the same founder in every post, the exact bottle with its exact label), use eroq's reference engines. [Eroq One](/models/eroq-one) builds a shot from up to ten reference pictures and edits an image from a plain-language instruction, for 20 credits / image. [Krea2](/models/eroq-krea2) and [Eroq Uncensored](/models/eroq-uncensored) hold a face from up to four references and edit in place. A workflow that plays to both sides: find the composition on a new model, then rebuild the keeper on Eroq One with your references.

Two more gaps. For anime and illustration with a negative prompt, [Anima](/models/eroq-anime) is the specialist. And when a provider turns a prompt down, eroq's own engines (Eroq Uncensored, Eroq One and Anima) are uncensored within the acceptable-use policy and fully private, because eroq runs them itself. One rule holds on every engine without exception: a prompt involving minors is refused before a single credit is charged.

## What a picture costs

Prices are flat per image. At the entry pack, $10 for 1,000 credits (about a cent a credit), a Sunburst image comes to about $0.31, a Seedream 5.0 Pro image to about $0.36 and a Nano Banana Pro image to about $1.24, the priciest image model on eroq. The same 124 credits buy one Nano Banana Pro take or 4 Sunburst takes, which is the whole case for drafting cheap and finishing expensive.

Subscriptions bring the per-credit price down: Hobby is $15 a month for 1,800 credits, Creator $49 for 6,000, Studio $149 for 20,000. The full table is on [/pricing](/pricing), and [what AI images cost in 2026](/blog/ai-image-generation-cost-2026) works out budgets model by model.

## From still to clip

Every image in your library has an **Animate** button. It opens the video studio with the picture as the start frame of an image-to-video clip, and you choose the engine there: [Seedance 2.0 Mini](/models/seedance-2-0-mini) is 45 credits for five seconds at 480p on any account, [Kling 3.0](/models/kling-3) is 170 for five seconds with sound from the Hobby plan up, and eroq's own [Motion One](/models/eroq-motion-one) needs no plan either. One catch: a Private-mode render stays in your browser, so it never reaches the library or its Animate button. The [image-to-video walkthrough](/blog/animate-ai-images-to-video) covers engine choice and motion prompts.

## For developers

The seven are in the API at the same price and policy as the studio. `POST /v1/images/generations` with a model id answers `202` at once with one job per take; poll `GET /v1/images/generations/{id}` until `status` is `succeeded` (image in `data`) or `failed` (refunded).

```json
{
  "model": "nano-banana-2-1",
  "prompt": "A flat vector trail map of a coastal hiking route, dotted path along the cliffs, a lighthouse at the end, sea in deep teal, sand-colored land",
  "aspect": "16:9",
  "batch": 2
}
```

The ids: `gpt-image-2-5-sunburst`, `gpt-image-2-5-flare`, `nano-banana-pro`, `nano-banana-2-1`, `seedream-5-0-pro`, `flux-3-image`, `grok-imagine-2-0`. The MCP server and the CLI take them too; see [/docs/images](/docs/images), [/docs/mcp](/docs/mcp) and the [API walkthrough](/blog/gpt-image-nano-banana-seedream-api) with curl and JavaScript samples.

## One guide per model

- [GPT Image 2.5 Sunburst guide](/blog/gpt-image-2-5-sunburst-guide): brief-style prompts, in-image text, layout.
- [GPT Image 2.5 Flare guide](/blog/gpt-image-2-5-flare-guide): key art, posters, cinematic stills.
- [Nano Banana Pro guide](/blog/nano-banana-pro-guide): when the priciest model pays for itself.
- [Nano Banana 2.1 guide](/blog/nano-banana-2-1-guide): fast drafts in plain English.
- [Seedream 5.0 Pro guide](/blog/seedream-5-0-pro-guide): detail, texture, photoreal prompting.
- [Flux 3 Image guide](/blog/flux-3-image-guide): natural light and typography.
- [Grok Imagine 2.0 guide](/blog/grok-imagine-2-0-guide): fast, characterful exploration.
- Head to head: [GPT Image 2.5 vs Nano Banana Pro](/blog/gpt-image-2-5-vs-nano-banana-pro), [Sunburst vs Flare](/blog/gpt-image-2-5-sunburst-vs-flare) and [Nano Banana Pro vs 2.1](/blog/nano-banana-pro-vs-nano-banana-2-1).

The fastest way to form an opinion: open the [image studio](/studio/image), write one prompt and run it on two models.

## FAQ

### Which image models did eroq add in October 2026?

Seven text-to-image models: GPT Image 2.5 Sunburst and GPT Image 2.5 Flare from OpenAI, Nano Banana Pro and Nano Banana 2.1 from Google, Seedream 5.0 Pro from ByteDance, Flux 3 Image from Black Forest Labs and Grok Imagine 2.0 from xAI.

### Do I need a subscription to use GPT Image 2.5 or Nano Banana Pro?

No. Neither needs a plan: a pay-as-you-go account renders both with a credit pack. A subscription only lowers what each credit costs.

### Can the new models edit a photo or match my reference pictures?

Not on eroq: all seven are text-to-image. For a face, product or outfit that must match your pictures, use Eroq One (up to ten references and instruction editing) or Krea2 and Eroq Uncensored (up to four references and in-place editing).

### Which new model should I try first?

Nano Banana 2.1, typically about 8 seconds a render. Then run the same prompt on Sunburst if it involves text, or on Seedream 5.0 Pro if it involves texture.