GPT Image 2: review, free access, API and prompts

GPT Image 2 is OpenAI's model for generating and editing images, and in the API its identifier is gpt-image-2. It came out on 21 April 2026, and on 8 September 2026 GPT Image 2.5 was added. The model has no free API: OpenAI charges for tokens. From Russia you can call it through a gateway paid in rubles, and below we go through how to build good requests.

What GPT Image 2 is and how 2.5 differs

In ChatGPT these models are called Images, and in the API they carry ordinary IDs. The search "chat gpt image 2" points at the same model inside the chat. A summary from OpenAI's documentation as of 29 October 2026:

All three accept text and images, return an image, and work through v1/images/generations, v1/images/edits and Batch. That is what the model pages for gpt-image-2 and gpt-image-2.5-flare say. The image generation guide names the quality levels xhigh, max and auto for version 2.5, along with the rules for arbitrary sizes: both sides a multiple of 16, aspect ratios from 1:3 to 3:1, no more than 3840 pixels on a side. OpenAI treats anything above 2560x1440 as experimental. A transparent background is set with the background and output_format parameters.

At codexkey version 2.5 is a single model under the ID gpt-image-2-5, whereas OpenAI's documentation lists two separate ones. Which variant is switched on for us, check on the model page in the catalog.

How much GPT Image 2 costs and whether it is free

OpenAI counts tokens, not pictures. The price list as of 29 October 2026, in dollars per million tokens, is the same for gpt-image-2 and both 2.5 models:

The price of a picture depends on size and quality. OpenAI's guide gives an example for GPT Image 2: medium quality, a 1024x1024 square, $0.053. The same guide adds a caveat: equal per-token prices do not mean equal prices per image.

Here is how you can get it free:

• ChatGPT. According to OpenAI's announcement, Images 2.5 is rolling out to all users of ChatGPT, ChatGPT Work and Codex on computers and phones. OpenAI publishes plan limits separately, and we found no figures for the free plan.
• API. The limits table on the model pages lists only the Build, Launch and Grow tiers, with no free one. Before your first request OpenAI may ask you to verify your organization.

Russia is not among the countries OpenAI supports, and the details are in Codex in Russia 2026. Sites offering "gpt image 2 free" do not belong to OpenAI, and what stands behind them is unknown.

How to write prompts for GPT Image 2

OpenAI's prompting guide advises going in order: scene, subject, details, constraints, and naming the purpose right away (an ad, an interface mockup, an infographic). Put literal text in quotation marks, describe the font and placement separately, and when editing say outright what stays unchanged. OpenAI itself names exact placement and text sharpness as the weak spot.

Six templates for different jobs:

1. Infographic. "Infographic titled 'How a coffee machine works' on a light background, four steps left to right with arrows. Step labels: 'Water', 'Heat', 'Pressure', 'Cup'. Sans-serif font, no extra words." A quoted list of labels and a ban on extras lower the risk of invented words.
2. Banner. "Horizontal banner for a sneaker sale: white sneakers on a red background, a large caption '30% OFF' on the left, once. No watermarks and no stray text." The words "once" and the bans work as a check.
3. Product photo. "Photo of a ceramic mug on a light wooden table, morning light from the left, shallow depth of field, a view from above at 45 degrees. A real photograph, not a render."
4. Edit. "Repaint the wall in the photo light green. Do not change the furniture, floor, shadows or lighting." Repeat the list of what stays unchanged in every iteration, or the picture drifts.
5. Interface mockup. "Screen of a food delivery mobile app: search on top, three dish cards below with a price and an 'Add to cart' button, a bottom menu of four icons. Light theme, rounded corners."
6. Cyrillic test. "A greeting card with the Russian caption 'Happy birthday, Anna!' in green letters in the center, nothing else written." Write the caption into the prompt in Russian, in Cyrillic letters, because that is what the test checks. Run it five times: the share of clean results shows whether you can put Russian text on a production line.

On quality OpenAI advises this: start with low when you send many requests, take medium for small text and portraits, and high for dense captions. The public codexkey documentation has no quality parameter, so you steer the result through resolution and size.

GPT Image 2 or Nano Banana Pro: a comparison by criteria

There is no single winner. We built the table from vendor documentation, and it holds no measurements of our own.

Here is how we choose. Mockups, signage and edits with a hard "do not change" list go to GPT Image 2 or 2.5. Infographics with current data and frames built from many references go to Nano Banana Pro. We check Russian text with the test from template 6 on both models. A general walk through the scenarios is in GPT Image and Nano Banana: how to choose an image API. Free and cheap routes for Google's model are collected in Nano Banana for free, and an open alternative under a non-commercial license is covered in the Qwen Image 2.1 review.

How to call GPT Image 2 through the codexkey API

You need a sk_cdx_... key; you can pay by a Russian bank card, SBP (the Russian Faster Payments System) or cryptocurrency. No OpenAI account is required. A request from the documentation:

curl https://api.codexkey.online/v1/images/generations \
  -H "Authorization: Bearer sk_cdx_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "gpt-image-2", "prompt": "a red apple on a white table", "resolution": "1K", "size": "1024x1024"}'

The same in Python, saving the file:

from pathlib import Path
from urllib.parse import urlparse

import requests

API_KEY = "sk_cdx_YOUR_KEY"

response = requests.post(
    "https://api.codexkey.online/v1/images/generations",
    headers={"Authorization": f"Bearer {API_KEY}"},
    json={
        "model": "gpt-image-2-5",
        "prompt": "A greeting card with the Russian caption 'Happy birthday, Anna!' in green letters in the center",
        "resolution": "1K",
        "size": "1024x1024",
    },
    timeout=300,
)
response.raise_for_status()

image_url = response.json()["data"][0]["url"]
suffix = Path(urlparse(image_url).path).suffix or ".png"
Path(f"card{suffix}").write_bytes(requests.get(image_url, timeout=60).content)

The link in data[].url is temporary, so save the file at once. The resolution field accepts 1K, 2K and 4K, and the values allowed for a model are listed on its page. Editing from a reference goes to POST /v1/images/edits, with details in the images and video reference. Money is charged only for a successful result.

A cheap route exists too: gpt-image-2 (in 1K and 2K) and gpt-image-2-5 are also sold in the Cheap channel. It costs less, but generation may fail more often. A key in cheap mode gets a 503 model_unavailable error without being charged when a generation fails, and 4K on gpt-image-2 always goes through Premium. Channels are explained in the Premium and Cheap guide. Look up image prices on the model page, because the public price list carries no figures.

Limits and errors

Three things are worth knowing in advance. First, according to OpenAI's model pages, gpt-image-2 and both 2.5 models work only through the image endpoints, and Chat Completions and Responses are not supported for them. Second, moderation. OpenAI may reject a request or a result with a moderation_blocked error and say at which stage it happened, on input or on output. Third, the link from data[].url is temporary, so move the file into your own storage right away. Gateway error codes are collected in the error reference.

FAQ

What is GPT Image 2?

OpenAI's model for generating and editing images, released on 21 April 2026. ChatGPT calls it Images 2.0, and in the API it is gpt-image-2. At codexkey it is available under the same ID.

Is there a way to use GPT Image 2 for free?

The OpenAI API has no free tier, and you pay for tokens. In ChatGPT images are available with limits, but Russia is not on the list of supported countries. Sites offering "gpt image 2 free" do not belong to OpenAI, and their terms cannot be checked.

What is GPT Image 2.5 and how does it differ from 2?

It is the update of 8 September 2026. In the OpenAI API it consists of two models: gpt-image-2.5-flare for fast everyday generation and gpt-image-2.5-sunburst for precise editing. It adds quality levels up to max and arbitrary sizes. At codexkey the version is called gpt-image-2-5.

How do I call GPT Image 2 through the API from Russia?

Get a codexkey key and send a POST to https://api.codexkey.online/v1/images/generations with the fields model, prompt, resolution and size. The link to the result arrives in data[].url. An example request is earlier in this article.

GPT Image 2 or Nano Banana Pro: which to choose?

GPT Image fits mockups and edits with a list of things to keep, and Nano Banana Pro fits infographics with current data and assembly from many references. Test both models on 15 of your own tasks first and compare the share of usable results.

How much does GPT Image 2 cost?

At OpenAI it is $30 per million output tokens, and the example for medium quality and a 1024x1024 square comes to about $0.053. At codexkey the price depends on the model and resolution, so check the model page in the catalog.