Image generation in ChatGPT is metered separately from text, on a per-plan allowance that resets on a cycle. OpenAI publishes current figures with its plan details rather than in the API reference, and they have changed several times, so checking directly is more reliable than any number quoted elsewhere. Through the API there is no plan allowance at all: you pay per image on dedicated models, which OpenAI lists as GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare in its model catalogue.
The number matters less than understanding what consumes an allowance, because most people burn through theirs on iterations that better prompting would have avoided.
Key points
- Image generation is metered separately from messages. A chat allowance and an image allowance are different things.
- Every attempt counts, including the bad ones. A generation you immediately discard consumed the same quota as one you keep.
- Edits usually count as generations. Asking for a change to an image typically produces a new one rather than modifying the existing file.
- The API charges per image rather than rationing, on dedicated image models separate from the text ones.
- Prompt quality is the main lever. Most quota is lost to iterations that a more specific first prompt would have prevented.
Why images are metered separately
The reason is computational rather than commercial, and understanding it explains the shape of the limit.
Generating an image is a fundamentally different operation from producing text. A text response streams token by token and can stop early. An image is produced as a whole through a process that runs to completion regardless of whether the result is any good. There is no partial output and no early exit.
That makes each generation a fixed, substantial unit of compute, which is why it is counted in images rather than folded into a general message allowance. It also explains why the allowances are small in absolute terms compared with message limits. A hundred images is a great deal more compute than a hundred messages.
OpenAI’s model catalogue reflects the same separation on the API side. Image work runs on dedicated models rather than on the general text models, with Sunburst positioned for the more capable generation and editing work and Flare for faster everyday output. Text models such as GPT-6 Astra accept images as input but do not produce them, which is a distinction worth knowing before designing anything around it.
What actually consumes your allowance
This is where people lose quota without realising, and all four causes are avoidable.
Discarded attempts count. The system meters generations, not keepers. Four attempts to get a usable image consumed four, and the three you rejected cost exactly the same as the one you kept.
Edits are usually new generations. Asking to change the lighting, remove an object or adjust the composition typically runs the model again rather than modifying pixels in place. A conversational editing session can consume a surprising amount of allowance quickly.
Variations count individually. Requesting several options at once produces several generations and is metered accordingly.
Regenerating after a small prompt tweak counts fully. There is no discount for a request that is almost identical to the previous one.
| Action | Counts as a generation | Note |
|---|---|---|
| First attempt | Yes | The only one most people expect to be counted |
| Rejected result | Yes | Metering happens on production, not on approval |
| Edit request | Usually | Typically reruns the model rather than altering the file |
| Variation request | Yes, per variation | Several options means several generations |
| Minor prompt tweak | Yes, in full | No discount for a near-identical request |
| Crop or colour fix outside the tool | No | Costs nothing against your allowance |
The pattern is that everything which produces an image costs an image, and there is no cheaper category for minor changes. Anyone working through many small adjustments will exhaust an allowance far faster than someone producing a few well-specified images.
This is genuinely different from how text works, and the mismatch causes most of the surprise. With text, an unsatisfactory answer costs little and refining through conversation is the natural and efficient way to work. Carrying that habit into image generation is expensive, because each turn of the conversation is a full generation rather than a cheap adjustment. The efficient approach inverts: think longer, ask once.
A realistic session, and where the quota goes
Walking through a typical session makes the arithmetic concrete, because people rarely count as they go.
Someone needs an image for a presentation slide. They type a short description and generate. The result is close but the framing is wrong, so they ask for it wider. That is a second generation. The lighting now looks flat, so they ask for something more dramatic. Third. The subject has drifted slightly from what they wanted, so they restate the description more fully. Fourth. This one is good but they would like to see an alternative. Fifth.
Five generations for one slide, and four of them existed only because the first prompt was three lines long instead of six. Nothing in that session was unreasonable, and it is how almost everyone works by default.
Now compare a session where the prompt is written out first, with subject, framing, lighting, mood and exclusions stated explicitly. The first result is usually close enough to use, occasionally needing one adjustment. One or two generations instead of five.
The difference is not skill with the tool. It is deciding what you want before asking, which costs two minutes and saves most of an allowance across a working week.
What to know before deciding
Several practical points follow from how the metering works.
Describe the image, not the subject. “A photograph of a rain-soaked street at night, low angle, wet asphalt reflecting neon, shallow depth of field” produces something specific. “A city at night” produces a generic result you will then iterate on, which costs quota.
Change one thing at a time. When you do iterate, adjusting a single element lets you attribute the difference. Rewriting the whole prompt each round means you never learn which word mattered, and you spend more attempts.
Say what you do not want. Negative direction is frequently more efficient than piling on positive description, particularly when successive attempts keep drifting the same way.
Plan the aspect ratio and use before generating. Discovering afterwards that you needed a vertical crop is an avoidable regeneration.
Check commercial terms before using output in paid work. Rights over generated images and permitted commercial use vary between providers and are updated periodically, so confirm the current position directly rather than relying on what was true last year.
The API is the answer to volume. If you routinely exhaust a plan allowance, per-image pricing on the dedicated models is a different economic model rather than a bigger version of the same one.
Getting more from each generation
- Write the prompt before opening the tool. Composing in the interface encourages generating early and iterating expensively.
- Include composition, lighting and framing explicitly. These are the three things people most often leave implicit and then regenerate to fix.
- Use a reference image for style where the tool supports it, since some looks are far easier to show than to describe.
- Accept a near-miss and edit elsewhere. A small crop or colour adjustment in an image editor costs no quota at all.
- Batch your thinking, not your requests. Deciding on three distinct images and specifying each properly beats generating twelve and picking the least wrong one.
Plan allowance versus paying per image
The two access routes have genuinely different economics, and knowing which suits you avoids both wasted quota and unnecessary spending.
A subscription allowance is predictable and capped. You know what you can produce in a period and you cannot accidentally spend more, which suits occasional and creative use where the volume is small and unpredictable. The trade is that unused allowance does not carry over and a busy week can leave you waiting.
Per-image API pricing is the opposite. There is no ceiling to hit and no waiting, and the cost tracks exactly what you produced. That suits repeated production work where volume is known and the images feed into something else, such as generating variants for a catalogue or illustrations for a documentation set.
The signal that you should move is simple: if you regularly reach a plan limit and the work is not discretionary, per-image pricing is the appropriate model rather than a larger subscription. The reverse also holds. If you exhaust an allowance occasionally and mostly through experimentation, that is a prompting problem rather than a plan problem, and upgrading treats a symptom.
One further consideration applies to teams. Plan allowances attach to individual accounts, so a team relying on image generation ends up rationing across people rather than across need, and whoever has quota left is not necessarily whoever has the most important task. Teams doing this seriously tend to move to the API for exactly that reason.
Decision framework
Five questions before you start generating.
- Do you know what the image needs to show? Vague intent is the single largest cause of wasted attempts.
- Have you specified composition and lighting? These determine whether the first result is close, and closeness decides how many attempts follow.
- Is this a one-off or a repeated need? Repeated volume belongs on the API where you pay per image rather than exhausting a plan.
- Could an edit outside the tool finish this? Cropping, colour and small removals are often faster and cost nothing.
- Do you need generation at all? For a diagram, a chart or anything requiring accurate text, a different tool usually produces a better result.
Knowing which tool fits which task, and how to specify a request so the first attempt lands close, is a transferable skill rather than a product feature. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
Your next step
Before your next session, write the full prompt in a text editor first: subject, composition, lighting, mood, and what you specifically do not want. Then generate once and see how close it lands.
Most people find the first result is far better than what they usually get, because the usual approach is a short prompt followed by four corrective attempts. The prompt takes two minutes and saves most of an allowance.