GPT Image 2.5 Flare and Sunburst have an Effort setting with five levels: Low, Medium, High, Extra high, and Max. Effort controls how much detail the model spends on each image. More effort means sharper fine detail, longer generation times, and a higher price.
It's easy to leave it on whatever it defaults to. So we ran the same five prompts and two edits through both models at every level to see what you actually get.
The short version: Medium often looks fine as a thumbnail and falls apart up close. Below High, skin often turns plastic and lighting looks over-processed. High is where skin and fine detail hold together most reliably, which is why High is now the default on Yapper. Extra high and Max add more micro-detail on top, at roughly two and four times the cost of High.
How We Tested
- Models: GPT Image 2.5 Flare and GPT Image 2.5 Sunburst
- Settings: 2k resolution (2304×1728), 4:3, no reference images
- Effort levels: Low, Medium, High, Extra high, and Max
- Prompts: a close-up portrait, a K-pop stage close-up, a group shot with hands, a neon street with signage, and a mechanical watch
- Edits: a wardrobe and lighting change, and a hair color and stage lighting change
Every image is a single generation, and each effort level is a separate generation, so compositions differ from one level to the next. Look at texture, lighting, and edges rather than framing.
Cost and Speed at Each Effort Level
| Effort | Credits at 1080 | Credits at 2k | Typical Sunburst time at 2k |
|---|---|---|---|
| Low | 2 | 3 | ~25 seconds |
| Medium | 4 | 5 | ~40 seconds |
| High (default) | 12 | 17 | ~55 seconds |
| Extra high | 21 | 30 | ~70 seconds |
| Max | 46 | 67 | ~2 minutes |
Credits are for a text-to-image generation with no references. Flare is usually faster than Sunburst at the same effort.
What Effort Actually Changes
GPT Image 2.5 builds each image out of output tokens, and effort sets how many it spends. More tokens give the model more room for fine detail, and they're also what you pay for. Here's what each level spends on a single 1024×1024 image, according to OpenAI's calculator:
| Effort | Output tokens |
|---|---|
| Low | 196 |
| Medium | 439 |
| High | 1,756 |
| Extra high | 3,122 |
| Max | 7,024 |
The jump from Medium to High is the biggest step on the scale: High spends four times as many tokens. That lines up with what we saw in the tests below, where High is usually the first level that holds skin and intricate detail together. It's also why High costs about three times as much as Medium.
Portraits Need at Least High
In this portrait, effort showed up in the face first. At Low and Medium the skin looks artificial: wrinkles look etched in with harsh contrast, the light has a flat, over-processed look, and the mud on the coat reads as painted on. High is the first level where the skin and light look like a real photograph. Extra high pushed the wrinkles almost too far, and Max looked the most natural of the five.
Not Every Prompt Shows a Gap
We expected a sweaty stage close-up to be a hard test for skin, but this one held up at every level. Low and Medium already have believable skin, sweat sheen, and stage lighting. High and above are only slightly crisper, mostly in the hair strands and the highlights on her shoulder.
The bigger problem had nothing to do with effort. Sunburst drew the headset microphone as a bead with a spike sticking out of it at every level, Max included. Flare drew it correctly, with the boom running back to her ear, at every level. More effort doesn't fix something a model misunderstands, but switching models can.
So effort isn't a guarantee either way. Some prompts fall apart below High, and some look fine at Low. High is the default because it holds up across the widest range of prompts.
Hands Got the Basics Right
Hands used to be the classic AI failure. GPT Image 2.5 got finger counts and grips right at every effort level we tried, and even Medium wrote a readable chalkboard sign. The difference is the overall look. Low and Medium come out contrasty and over-processed, like a heavy filter. High and Max have soft, natural window light and read like a real candid photo.
Signs Look About the Same
Text held up better than we expected. The signs we asked for by name, and the other big signs, are readable at every level, Medium included. Small signs in the distance turn into shapes that only look like writing at every level, High included. High did get a few more of the larger background signs right, but if you need exact text in an image, effort alone won't guarantee it.
Fine Detail Is the Biggest Difference
This is the prompt that separates the effort levels. The slider below shows 100% crops from the center of each Flare image:
At Low, parts of the movement melt into shapes that don't look like gears. Medium is busier but still smears in places. High is the first level where every gear, screw, and jewel reads as a real part. Extra high and Max add crisper engraving and finer brushed-metal texture.
Sunburst showed the same pattern. At Medium, the lettering on the dial came out garbled, with "SWISS", "MADS", and "MADE" stacked on top of each other. None of the higher-effort Sunburst runs garbled their dial text, and the gears are cleaner too:
Edits: The Source Carries the Detail
Editing works differently. When you give the model a photo to change, it keeps most of the detail from that photo, so the effort level matters less than it does for new images. We started from two of our High results and asked each model for one change.
Every effort level on both models made the change and kept the face, pose, and framing. The differences are in the parts the model has to redraw. At Low and Medium, relit skin and the new coat and hair come out a little softer, and sweat highlights lose some crispness. High matches the texture of the original photo. Extra high and Max add very little on top for an edit like this.
If you're making a quick change to a good photo, a lower effort can hold up better than it would for a new image. For anything with faces or fine texture, High is still the safe choice.
More Comparisons
Below are more comparisons you can go through yourself to spot differences between the models and effort levels. Each result is a single generation, so some differences may come from normal run-to-run randomness rather than from the model or the effort level itself. Every slider lets you pick any model and effort on either side, and the time shown is how long each image took to generate.
Styles and Design
Seven illustration styles, a website mockup, and a metro map, each run through both models at every effort level. Each slider opens on Sunburst High against Flare High; pick any model and effort on either side.
Across the seven styles, every piece of requested text came out spelled correctly at Low and at Max on both models: the diner sign, the picture book title, all four comic speech bubbles and the caption, and both lines of the poster. The two design tests below are where effort made a visible difference.
Anime
2D Cartoon
Stylized 3D
Watercolor
Comic with Dialogue
Pixel Art
Vector Poster
Website UI Design
Both models spelled every requested headline, button, product name, and price correctly at Low, High, and Max, and followed the layout from the prompt.
Metro Map with Labels
This was the clearest effort test outside photography. At Low, Sunburst's labels overlap lines and a few come out garbled. High and Max are much cleaner. Flare still garbled some station names at High, and was clean only at Max.
Detailed Photo Prompts
Six longer, more detailed photo prompts, each run on Sunburst and Flare at every effort level, 2k, 4:3. Times are how long fal spent generating each image. Pick any model and effort on either side.
Portrait
Stage Close-Up
Hands and a Group Scene
Packaging Text
Fine Mechanical Detail
Dense Scene with Signs
More Edits
Two detailed edits, run on both models at every effort level.
Rainy Dusk
New Label
Which Effort Should You Use?
- Low: quick composition tests you'll throw away. Don't judge a prompt by its Low result.
- Medium: rough drafts, like storyboards or layout checks. It can look fine on some prompts, but it's the first level to lose intricate detail and realistic skin.
- High: the default and the right choice for most images. It's the level where skin and fine detail hold up most reliably.
- Extra high: hero images, product shots, posters, and anything you'll view full screen or print. It costs about twice as much as High for a smaller step up.
- Max: the most detailed results in most of our tests, especially in fine texture. It costs more than twice Extra high and takes around two minutes, so save it for final images where every detail matters.
If a GPT Image 2.5 result looks soft or distorted, check the effort setting before you rewrite the prompt. Moving from Medium to High fixes a lot of what looks like a prompt problem.
You can change effort from the Effort control next to the model picker when Flare or Sunburst is selected. Yapper Assistant uses High unless you ask for something else.
Every Result
Every image from the detailed prompt tests, side by side. Switch models with the tabs and click any image to open it full size.
Generate with Flare or Sunburst and switch effort levels from the composer to compare them on your own prompts.