You already use FLUX.2 Klein for cheap drafts and edits, and now there is a FLUX 3 Klein next to it. Is it worth switching, and which of the three (FLUX 3 Klein, Klein 9B, Klein 4B) should take which job? The same applies to NSFW work, where all three are unfiltered on SpicyAPI but deliver different amounts. Short answer first, because the details change the price more than the picture.
Use Klein 4B for volume, Klein 9B when you want a slightly cleaner single image, and FLUX 3 Klein when the prompt contains numbers, the edit must keep a scene intact, or you need more than three reference images. FLUX 3 Klein delivered the first three intensity levels as written, where FLUX.2 Klein toned some down, but it costs more per image and its 2K tier is four times its 1K tier.
We ran one prompt set through all three on SpicyAPI. The FLUX.2 Klein images come from our 8 October blind-test run and the FLUX 3 Klein images from 22 October, one image per model per prompt. That is enough to see habits, not enough to rank them: we did not run a blind score for FLUX 3 Klein, so no quality number is quoted for it.
The short version
| Klein 4B | Klein 9B | FLUX 3 Klein | |
|---|---|---|---|
| Released | 15 Jan 2026 | 15 Jan 2026 | 15 Oct 2026 |
| Size | 4B | 9B | 5.6B image model |
| Text-to-image, 1K / 2K | $0.0093 / $0.0099 | $0.011 / $0.0122 | $0.012 / $0.048 |
| Edit price | $0.0126 flat | $0.0163 flat | from $0.0192 at 1K, more with extra references |
| Reference images per edit | up to 3 | up to 3 | up to 10 |
| Aspect ratios | 9 | 9 | 15, plus custom size up to 1024 |
| Output format | JPEG | JPEG | JPEG, PNG or WebP |
| Prompt expansion switch | no | no | yes, off by default |
| Typical time in our runs | 15.7 s median | 22.2 s median | 9 to 14 s in six runs |
Prices are the current SpicyAPI list prices; the FLUX 3 Klein, Klein 9B and Klein 4B pages always show today's number. The timings are not a controlled race: the FLUX.2 figures are medians over 14 prompts on 8 October, the FLUX 3 figure is six runs on 22 October.
Black Forest Labs reports FLUX 3 Klein ahead of FLUX.2 Klein 9B by 91 Elo in text-to-image and 70 in editing, according to a third-party review that quotes those figures. Treat that as the publisher's claim; our own pictures below tell a more mixed story.
Same prompt, three models
Each row of images below reads left to right: Klein 4B, Klein 9B, FLUX 3 Klein.
Counting
The prompt asked for a person holding up both hands, three raised fingers on the left and two on the right.
Left to right: Klein 4B, Klein 9B, FLUX 3 Klein. Only FLUX 3 Klein showed three and two.
Klein 4B did not show a count at all and Klein 9B showed three and one. FLUX 3 Klein got three and two on the first try. A second prompt asked for exactly five balloons, two red, two yellow and one blue, tied to a park bench with a dog asleep underneath.
Left to right: Klein 4B, Klein 9B, FLUX 3 Klein. Klein 9B added a green balloon.
Klein 4B and FLUX 3 Klein both drew two red, two yellow and one blue. Klein 9B added a sixth, green balloon. FLUX 3 Klein's strings drift in the air rather than being tied to the bench, so "tied" is the part it missed. If your work depends on counts, FLUX 3 Klein was the steadier of the three here, but run a few seeds and check.
Edits that keep the scene
For the edit test we sent one photo of a red lakeside cabin and asked for the morning after a heavy snowfall, with the lake frozen over and the cabin and camera angle unchanged.
Left to right: Klein 4B, Klein 9B, FLUX 3 Klein. Klein 9B turned the red cabin brown.
FLUX 3 Klein kept the cabin red and showed the lake frozen with visible ice. Klein 4B kept the red but covered the lake in snow, so "frozen over" is less clear. Klein 9B changed the cabin to brown, something it was told to keep.
The opposite happened with a restyle. We asked for a loose watercolor painting while keeping every building, awning and bicycle in place.
Left to right: Klein 4B, Klein 9B, FLUX 3 Klein. FLUX 3 Klein stayed closest to the photo and painted the least.
Klein 4B and 9B produced real watercolor looks; FLUX 3 Klein stayed closest to the source, with a watercolor sky and a photographic street. That is good if you want faithful edits and a drawback if you want a strong style change. Say how strong a change you want in the prompt.
Faces and groups
The prompt asked for friends of different ages and ethnicities laughing around a garden dinner table.
Left to right: Klein 4B, Klein 9B, FLUX 3 Klein.
FLUX 3 Klein's table spans clearly different ages, including a grey-haired man; Klein 4B and 9B drew younger groups, with Klein 9B showing more variety in skin tone.
Across the other prompts we looked at, the differences were smaller. A requested wine glass filled to a bulging brim stayed below the rim on all three. "GOOD MORNING" printed on a mug was spelled right by all three, with Klein 4B printing it far larger than asked. On a macro watch photo FLUX 3 Klein's movement looked the most detailed to us. In an anime scene Klein 9B's jacket patch carried garbled letters.
Speed and price: where FLUX 3 Klein is not the cheap one
On a 1K text-to-image, FLUX 3 Klein is the dearest of the three: $0.012 against $0.011 for 9B and $0.0093 for 4B. At 2K it jumps to $0.048, while 4B and 9B stay near a cent. For 1K drafts the gap is small. For 2K batches Klein 9B is about four times cheaper. Edits are the same story: $0.0126 for 4B and $0.0163 for 9B, flat for one to three references, against $0.0192 and up for FLUX 3 Klein.
On speed, FLUX 3 Klein finished in 9 to 14 seconds in six runs, with the 4B and 9B medians at 15.7 and 22.2 seconds on a different day. Add queue time on busy days. Failed requests are not charged on any of them.
Prompts: what to write differently
- Reference images. FLUX 3 Klein names references as
<ref_image_1>,<ref_image_2>and so on, up to ten. FLUX.2 Klein expects "image 1", "image 2" and "image 3", up to three. - Prompt expansion. FLUX 3 Klein has an
enable_prompt_expansionswitch, off by default. It rewrites your wording, so use it on purpose. FLUX.2 Klein has no such switch; write the full scene yourself. - Be literal about counts and colors on FLUX.2 Klein. In our runs hands and counts drifted more often there, so plan to run a few seeds.
- Say how strong an edit should be. FLUX 3 Klein stayed close to the source in our restyle test; if you want a strong style change, say so in the prompt.
- No negative prompt on any of them. Describe what you want to see.
How far each one goes
All three are unfiltered on SpicyAPI: SpicyAPI adds no filter on prompts, source images or results, and none refused anything in our tests. The difference is how much the model itself delivers. We test at five intensity levels (Mild, Moderate, Strong, Very strong, Maximum; the test prompts are not published) and report delivered, toned down or refused.
- FLUX 3 Klein: Mild to Strong delivered as written on text-to-image and edit. Very strong edits delivered; Very strong text-to-image came back as asked in at most one run of three, and Maximum was toned down. Turning prompt expansion on made Very strong text-to-image come back as asked three times out of three.
- Klein 9B: Mild and Moderate delivered on both endpoints. Strong text-to-image was toned down in all three runs, while the Strong edit delivered; Very strong was toned down on both endpoints.
- Klein 4B: the more conservative of the two. Moderate text-to-image was mostly toned down and Strong and Very strong text-to-image were toned down; edits delivered up to Strong and toned down at Very strong.
The FLUX.2 Klein results are from 8 October and the FLUX 3 Klein results from 16 and 22 October, so read them as habits rather than a controlled head-to-head. The full write-ups are in our FLUX.2 Klein tests, the FLUX 3 Klein review and the reviews of Klein 9B and Klein 4B. Adults only in every prompt and every source photo, and never upload a real person's likeness without consent.
Which one should you use?
| Your job | Pick | Why |
|---|---|---|
| Hundreds of 1K drafts or thumbnails | Klein 4B | Cheapest, quick and steady |
| One clean single image, no big edit | Klein 9B | Image quality 73.4 against 4B's 69.9 in our 8 October blind test |
| Prompts with counts, hands or exact colors | FLUX 3 Klein | Got the finger count and balloon colors right in our runs |
| Edits that must keep the scene, the person or the color | FLUX 3 Klein | Kept the red cabin; 9B changed it |
| Four to ten reference images in one edit | FLUX 3 Klein | The only one of the three that takes more than three |
| Panoramas, 4:5, 3:1 or a custom size | FLUX 3 Klein | Fifteen ratios plus custom width and height |
| 2K batches on a budget | Klein 9B or 4B | $0.0099 to $0.0122 against $0.048 |
| Strong painterly restyles | Klein 4B or 9B | FLUX 3 Klein stayed close to the photo |
A sensible default: draft on Klein 4B, then re-run the keepers on FLUX 3 Klein when numbers or edit fidelity matter. If you need a mask or a transparent background, neither family does that; see Qwen Image vs FLUX.
FAQ
Is FLUX 3 Klein better than FLUX.2 Klein?
For counts, edit fidelity and reference images, yes in our runs. For price per 1K image and strong restyles, FLUX.2 Klein can be the better fit. We ran one image per prompt, so look at your own prompts before moving a whole batch.
Is FLUX 3 Klein or FLUX.2 Klein better for NSFW?
SpicyAPI adds no filter on either. FLUX 3 Klein delivered Mild to Strong as written on text-to-image and edit in our tests, while FLUX.2 Klein toned down text-to-image at Strong; see the section above for the five-level breakdown.
Klein 4B or 9B?
4B for volume and speed (15.7 s median, $0.0093 at 1K); 9B when you want a bit more image quality for about 17 cents more per hundred 1K images. Both have the same two endpoints and the same three-reference limit.
How many reference images can I use?
Up to three on FLUX.2 Klein and up to ten on FLUX 3 Klein.
Does FLUX 3 Klein cost more?
Yes: $0.012 at 1K and $0.048 at 2K for text-to-image, against $0.0093 to $0.0122 for the FLUX.2 Klein models.
