GPT Image 2.6 團隊The short answer: GPT Image 2.6 gives you two dials and one habit. Quote text that must be exact, raise complexity when the layout is dense or many parts must agree, and raise quality when the layout is right but small type blurs. Then change one dial per run, because two changes hide the cause. Both variants plan at high unless you set it, so the default is already the top planning level.
Quotes are for characters, not for emphasis
A quoted string is the strongest signal you can send about exact text. OpenAI's guidance is to put required wording in quotes, then describe where it sits and how it is set.
Those are three separate facts: the string, its position, and its type. Each one removes a guess.
A matte ceramic mug on a walnut desk, morning window light from the left. Label text "MORNING BLOOM" in small caps, centered on the mug. No other text.
Remove the quotes and the same words become a description of a mood, not a required string. Nothing then tells the model to keep them exact.
Position and type carry real weight. "Centered on the mug" keeps the string off the corner. "Small caps" keeps a display font from swallowing the frame.
Spell unusual brand names letter by letter when spelling is the point. Add "no other text" too, then read the result for stray lettering.
Quotes decide which characters appear. They do not decide sharpness, and that is the next rule.
Density sets the complexity budget
Complexity buys planning, and planning pays off when many parts must agree. OpenAI defines the field as how much the model plans before it renders.
Dense labels, charts, and object relationships are where that planning shows up. A poster with six callouts asks for more arranging than a single portrait.
The setting has a limit worth knowing. When a prompt already fixes layout, text, and style, OpenAI notes that higher complexity may make little visible difference. There is nothing left to decide.
| What the image must hold | Start at | Check first |
|---|---|---|
| One subject, open composition | high (the default) |
where props and text landed |
| Dense labels, diagram, chart | high |
label placement and connected objects |
| Layout, text and style already fixed | any level | whether the output changed at all |
Complexity has no separate fee on the API. What it can cost is time, since OpenAI says higher complexity may add latency. It can also bring changes you did not ask for.
Quality and complexity fail in different places
Read the failure first, then pick a dial. The two settings break in different spots, so the symptom tells you which one to move.
OpenAI's own test is a clean split. If the layout is correct but small text is hard to read, compare quality settings at the same complexity. If the structure is wrong, the problem sits upstream of rendering.
| What you see | Move | Hold fixed |
|---|---|---|
| Small type blurs or misspells | quality up |
complexity |
| Labels, object count or connections wrong | complexity up |
quality |
| An edit changes parts you wanted kept | complexity down, per OpenAI's focused-edit case |
quality |
The API takes six quality levels, from auto to max. This site offers three: low, medium, and high. Low and medium cost the same here and high adds 50%, so spend high on frames with small type. Credits rise with image size; the pricing page has the rates.
Move one dial per run
A run that moves two settings teaches you nothing. OpenAI's migration steps say to tune one setting at a time, and that is the whole discipline.
Keep the prompt, size, references, and model fixed. Change quality or complexity, then run the same request more than once. One output is a sample, not a result.
Track how often the image clears your bar, not how good the best one looked.
Then write each failure down in plain words, such as "third label misspelled" or "hands merged." A saved description beats a memory of a good image.
Plan less while the layout is open
Drop complexity when the composition is still a question. Low complexity suits focused work, and a draft is focused work.
OpenAI names small, targeted changes as the low-complexity case, such as a background or a sign. A first pass follows the same logic. You are testing an idea, not shipping the frame.
The costs run the other way too. High complexity can add latency and unwanted changes, which is risk without reward on a tight edit.
Once the layout is locked, go back to the default and spend quality where the small type lives. Switching to gpt-image-2.6-sunburst is the other lever for detail-heavy frames, and the API reference sets it out next to the token rates.
You can run both variants in the browser generator without an OpenAI account. Failed runs refund the credits, so a bad guess costs time rather than money.
FAQ
Do I need quotes around every line of text?
Only around text that must appear word for word. Quote headlines, labels, and legal lines. Leave quotes off words you use to describe a mood.
Is low complexity faster?
OpenAI says higher complexity may increase latency. That is the only speed guidance it gives, and it publishes no figure for either variant, so treat exact numbers you see elsewhere as guesses.
Does higher quality fix a wrong layout?
No. Quality governs rendering detail such as small lettering. A wrong object count or a misplaced label is a planning problem, so that one is complexity.
Does complexity cost extra?
No. Adjusting complexity has no separate fee on the API. Your bill moves with quality and size instead.
Which variant should I prompt?
Try gpt-image-2.6-flare first when response time matters, and gpt-image-2.6-sunburst when detail or layout accuracy decides the result. The comparison with 2.5 covers what stayed the same between releases.
Do these rules apply to edits?
Yes. The split holds on edits. Say what changes, list what must stay, and keep quality fixed while you test complexity.
Sources
- Image prompting guide — the 2.6 parameter table, the quoted-text rule, the quality and complexity test, and "tune one setting at a time."
- Image generation guide — the "Set image complexity" section, including the no-separate-fee line.
- Changelog — the November 4, 2026 release note for GPT-Image-2.6.
- API pricing — standard per-token rates for both 2.6 variants.
- gpt-image-2.6-sunburst and gpt-image-2.6-flare — model pages.
Revision log
- November 5, 2026 — Page published.