Long Prompts and Two-Stage Generation: Where the 32K Limit Comes From

​

The customer’s question: why does the API cap at 32K when ChatGPT does not?

Customer: Your prompt length is different from ChatGPT’s. Looks like it is capped at 32K. Us: There is a limit. Who has an image job with a 32K prompt? Customer: Plenty. Ad creatives, packaging dimensions… I will just call codex cli myself. Us: To confirm, a 32,000-token prompt? Customer: Yes. 32,000 in English is not long at all.

​

Where the 32K limit comes from

​

Why the ChatGPT web app looks unlimited

• Paste a long document into ChatGPT and the reader is the chat model, not the image model. Having read it, the chat model writes its own short image prompt and calls the image tool with that. The image model never sees the original material. Pastes over 10,000 characters are even turned into attachments automatically (OpenAI Help Center, help.openai.com), which makes it plain that the text is for the chat model.
• Through the API you talk to the image model directly. Nobody reads the material and makes choices for you in between. The limit belongs to the image model layer, and the web app never exposes the image model to it.
• The customer’s parting line, “I will just call codex cli myself”, is the right instinct: have a text model read the material and produce the image prompt. That is exactly what the web app does behind the scenes, and the two-stage pipeline below makes it runnable.

​

The two costs of a long prompt

​

Two stages: material into a text model, prompt into the image model

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APIYI_API_KEY"],
    base_url="https://api.apiyi.com/v1",
)

DISTILL_SYSTEM = """You are an ad-creative prompt engineer. Read all the material the user provides and output one prompt that can be sent directly to an image model.

Write in this order, one line per section:
1 Hard constraints: logo not distorted, packaging text verbatim, number of people, aspect ratio, background colour value
2 Purpose and placement
3 Subject and packaging elements that must be kept
4 Composition and negative space: product position, people position, framing, area reserved for copy
5 Visual: setting, palette, one identifiable key light, materials, photographic style
6 Do not

Rules:
- At most 2500 characters. Output only the prompt body, no explanation, no headings
- Do not invent anything absent from the material, especially brand names and packaging text
- Do not use vague quality words such as 8K, ultra HD, masterpiece, perfect
- Keep anything the material specifies explicitly, do not rewrite it"""

PROMPT_LIMIT = 32000  # OpenAI Images API prompt limit, in characters


def distill(brief: str, model: str = "gpt-5.6") -> str:
    resp = client.chat.completions.create(
        model=model,
        messages=[
            {"role": "system", "content": DISTILL_SYSTEM},
            {"role": "user", "content": brief},
        ],
    )
    prompt = resp.choices[0].message.content.strip()
    if len(prompt) > PROMPT_LIMIT:
        # rare; compress one more round if it happens
        prompt = distill(f"Compress the following prompt to under 2500 characters, keeping every hard constraint:\n\n{prompt}", model)
    return prompt


def generate(prompt: str, size: str = "1536x1024", quality: str = "high") -> bytes:
    import base64
    resp = client.images.generate(
        model="gpt-image-2.5-sunburst",
        prompt=prompt,
        size=size,
        quality=quality,
        timeout=600,
    )
    return base64.b64decode(resp.data[0].b64_json)


if __name__ == "__main__":
    brief = open("brief.md", encoding="utf-8").read()   # brand book + packaging spec + ad requirements, any length
    prompt = distill(brief)
    open("prompt.txt", "w", encoding="utf-8").write(prompt)   # persist; generation replays only this
    open("ad.png", "wb").write(generate(prompt))

​

When a long prompt is genuinely needed

• Multi-image editing: refer to reference images as “image 1 / image 2 / image 3” and say what to take from each. A few hundred characters.
• Text inside the image: signage, posters, packaging copy must be given verbatim rather than left to the model. Tens to hundreds of characters.
• A shared prefix for a series: the style, lighting and composition block common to a batch. Under a thousand characters.

​

Quick reference

• 32,000 characters is the official OpenAI Images API limit, counted in characters not tokens, identical for /generations and /edits; APIYI’s official relay does not tighten it.
• Characters, tokens and context window are three different things: 32K English is about 8K tokens, Chinese is more; a text model’s context window has nothing to do with the image model’s prompt limit.
• “No limit” on the web app is an illusion: the chat model reads the material and writes its own short prompt for the image tool; the image model never faces the 32K.
• Detailed is not long: text input is billed per token and every retry pays again; hundreds of competing requirements bury the hard constraints.
• Two stages: distill the material with a text model into a 1K to 3K character structured prompt, hard constraints first, persist it, and replay only the prompt for generation.

​

Related docs

• How to Get Satisfying Images
• Safety Rejections
• Advanced Image Generation: Workflow and Realism
• Image Prompt Doctor Skill
• Text-to-Image API Reference