Guide
GPT-6 Astra image generation: quality, how to use it and what it costs
7 min read
GPT-6 Astra is OpenAI's first GPT-6 model, and most of the coverage is about reasoning, coding and computer use. The question app marketers are asking is simpler. Can it make good images, and how do you get them?
This guide answers that with the facts as they stand a day after launch, including what is and is not confirmed.
Does GPT-6 Astra generate images?
Not on its own. OpenAI's model page lists Astra's input as text and image and its output as text only. When you ask Astra for a picture, it calls a separate image model through a tool and hands you the result.
That image model is GPT Image 2, released on April 21, 2026. ChatGPT's built-in image generation runs on it, and the API's image_generation tool selects it automatically when Astra is the model in the conversation.
So the quality of Astra's images is the quality of GPT Image 2. What Astra adds is the reasoning around it: understanding a vague brief, planning the shot, and iterating on edits across a conversation.
What is GPT-6 Astra?
Astra launched on September 3, 2026 as a model built for long-horizon reasoning, agentic software work and computer use. It has a context window of about 1.05 million tokens, a maximum output of 128,000 tokens and a knowledge cutoff of April 30, 2026.
It is rolling out in phases. Enterprises in OpenAI's Trusted Access Program and API developers came first, with ChatGPT Plus, Pro, Business and Enterprise access following over the coming weeks. There is no confirmed date for the Free plan.
None of the launch benchmarks measure image output. Treat the numbers above as a signal of how well Astra understands and plans, which is what it brings to image work.
How good is the image quality?
GPT Image 2 is an autoregressive model rather than a diffusion model, and it is the first in the series to put a reasoning step inside generation. In practice that shows up as strong instruction following, reliable on-image text and edits that change one thing while leaving the rest intact. It also renders up to four times faster than GPT Image 1.
The weak spots are documented. Earlier versions over-sharpened and leaned warm, which 1.5 largely fixed. Scenes with several faces still go wrong more often than single portraits, and text in scripts such as Arabic, Chinese and Hebrew is less reliable than Latin text. One early Astra reviewer also noted its visual design output can trail Anthropic's Claude even when the underlying work is strong.
For ad creative, the practical summary is this. Product shots, UI mockups, lifestyle stills with one or two people and images that need legible headline text are excellent. Crowd scenes and non-Latin typography need a second look.
How do you generate images with Astra in ChatGPT?
Once Astra is available on your plan, there is no separate mode to switch on. Describe the image in the composer and it generates.
- Select GPT-6 Astra in the model picker, then write one to three clear sentences: subject, setting, composition and style.
- Attach reference images if you want a specific product, face or layout preserved.
- Open a result in the expanded viewer. Focused view inspects one image, Canvas view shows every version side by side.
- Use Comment to point at a region and describe the change. Astra edits that area and leaves the rest alone.
- Multi-select several images to apply the same edit or style to all of them at once.
- Ask for text by giving the exact words in quotes and the placement. GPT Image 2 renders it as written.
Image generation counts toward your plan's usage limits, and OpenAI says availability and limits depend on plan and workspace settings. Expect Pro to allow far more images than Plus, and check the limits page for your tier.
How do you use it through the API?
Use the Responses API with gpt-6-astra as the model and add a tool of type image_generation. Astra decides when to call it, and because the conversation keeps state you can ask for edits in follow-up turns with high fidelity.
The tool picks the image model for you. Its parameters control the output.
Two details worth knowing. The input_fidelity setting that preserved reference detail on older models is not configurable for GPT Image 2, which handles it automatically. And Chat Completions does not expose the tool, so use Responses.
What does it cost?
Two meters run when Astra makes an image. Astra's own tokens for the reasoning and the conversation, then GPT Image 2's tokens for the picture. Astra is a premium model, so keep the conversation tight and let the image model do the work.
A single high-quality image lands in the tens of cents once Astra's overhead is included. For batches of draft concepts, run the Image API directly with GPT Image 1 Mini and bring Astra in only for the finals.
Prompting tips for ad-ready images
- Write the image the way you would brief a photographer. Who, where, doing what, shot how, in what light.
- Give exact on-image text in quotes and say where it goes. Do not paraphrase it.
- Use 1024x1536 portrait and leave headroom. A 9:16 story crop comes from the centre.
- Attach a product screenshot or packshot as a reference so the UI or the object stays accurate.
- Fix one thing per edit. Point at the region, describe the change, keep the rest.
- Ask for one or two people, not a crowd, and check hands and overlapping objects before you ship.
Stills or video for app ads?
Astra and GPT Image 2 give you excellent stills for App Store screenshots, static feed ads and thumbnails. They do not give you a moving, talking, native-looking video, and on TikTok and Reels that is what earns the install.
That is the job Viralway does. Write a five-line brief and it returns hooks, scripts, voices and finished vertical clips with captions the same day. Use Astra for the stills, Viralway for the video, and test both.
Key takeaways
- Astra outputs text only. Its images come from GPT Image 2 through a tool call.
- GPT Image 2 leads on text rendering, instruction following and precise edits. Multiple faces and non-Latin scripts are the weak spots.
- In ChatGPT, just describe the image. Use Comment and Multi-select for edits.
- In the API, use Responses with the image_generation tool. Chat Completions does not support it.
- You pay Astra tokens plus image tokens. Draft with GPT Image 1 Mini, finish with Astra.
Ready to put this into practice?
Send Viralway a brief and get platform-ready clips for Meta, TikTok and YouTube the same day.
Common questions
- Is GPT-6 Astra image generation free in ChatGPT?
- Not at launch. Astra is rolling out to Plus, Pro, Business and Enterprise first, and image generation counts toward each plan's usage limits. Free plan access has no confirmed date.
- Can Astra make 9:16 images for TikTok and Reels?
- Not natively. Supported sizes are 1024x1024, 1536x1024 and 1024x1536. Generate portrait and crop to 9:16, keeping the subject centred.
- Is Astra better than Midjourney or Nano Banana for images?
- It depends on the job. GPT Image 2 is the safer choice when the image needs exact text, a preserved product or a precise edit. Midjourney still leads for painterly style, and no published benchmark compares Astra's tool output head to head yet.
