Flux 3 vs Flux 2: a fact-based family comparison.

The comparison is asymmetric right now

A conventional version comparison assumes two products that perform the same task. The published FLUX record is broader: FLUX 3 Video is a per-second video API, while FLUX.2 includes still-image generation and editing products. Black Forest Labs has not attached separate public Image specifications or a Dev package to the FLUX 3 timeline as of August 10, so image benchmarks, checkpoint sizes, licenses, and image prices cannot be compared on equal terms.

That does not make comparison impossible; it changes what can be compared responsibly. Modality, access type, pricing unit, and deployability are confirmed operational parameters. Claims about future Image quality, prompt adherence, typography, editing fidelity, speed, or hardware requirements remain outside the published record.

Any table that fills those open cells with confident numbers is projecting from the family announcement or confusing FLUX 3 Video with a still-image product. Leaving a cell as “not published” is more informative than inserting a guess.

What FLUX.2 offers for image work today

FLUX.2 pro is a hosted still-image option with a published price of $0.03 per generated image and $0.045 per edit. That creates a useful image-cost reference alongside FLUX 3 Video's per-second economics, although the two modalities should not be compared as if one output unit represented the same creative job.

The wider FLUX.2 price ladder includes klein 4B at $0.014 per image, klein 9B at $0.015, flex at $0.05, and max at $0.07. A lower API price is not automatically the best operational choice. Teams should also evaluate output quality on their briefs, edit capability, latency, reliability, rate limits, and the cost of supporting several technical options.

For product design, the lesson is to separate internal optionality from interface complexity. A service can evaluate several technical paths while presenting users with one predictable generation workflow and one visible credit charge. The family comparison belongs in technical research, not in a required model-selection step before every prompt.

What FLUX 3 Video contributes to the comparison

FLUX 3 Video establishes that the family is more than an announcement. It has a production API and published usage prices: full rendering at $0.17 per second for HD and $0.29 per second for FHD, draft generation at $0.06 per second, and video extension at $0.43 per second for HD. Those are video economics, not image prices.

General availability also provides a useful release-process signal. Black Forest Labs moved from the July 23 multimodal announcement to video GA on August 4. That shows staged delivery, but it does not reveal a schedule for other products. Video infrastructure, evaluation, safety, and capacity differ from still-image inference and open-weight distribution.

A cross-modal comparison could eventually examine shared prompting or visual continuity. For the current table, using the Video release as evidence for an undocumented image capability would go beyond the confirmed product surface.

How to compare image systems fairly

Instruction following should be tested with briefs that contain spatial relationships, material constraints, and explicit crop requirements. Visual quality should be broken into observable categories such as anatomy, texture, repeated geometry, text rendering, reflection behavior, and lighting consistency. Edit performance should measure whether the system preserves named invariants while changing only the requested attributes.

Operational parameters matter just as much. The comparison should record input limits, supported ratios, output resolution and format, latency distribution, error rate, concurrency limits, price, content rules, and whether results are returned synchronously or through polling. For open weights, license, parameter count, precision, memory requirements, quantization options, and reference inference code become essential.

A fair evaluation saves the prompts and input conditions so the same brief can be tested across products without changing its visible constraints. That produces evidence grounded in identical tasks rather than unrelated showcase images selected for different strengths.

Which should you use now?

For still-image API research, FLUX.2 provides published price and access references across pro, klein, flex, and max. Preserve representative briefs and selected ratios so later family-level comparisons use the same inputs instead of relying on memory or vendor showcases.

For video generation, evaluate FLUX 3 Video against duration, resolution, draft, and extension needs. Its price is measured per second, so storyboard duration and iteration strategy have direct cost implications. Draft mode may reduce the expense of composition tests before a final-resolution run.

For local or self-hosted FLUX 3 development, wait for an official Dev package and license. A hosted video endpoint is not an open-weight release. Architecture work can still use provider interfaces and model registries without pretending to know an unpublished checkpoint format.

The durable conclusion is that version numbers do not erase product boundaries. Compare video with video, image with image, hosted access with hosted access, and open weights with open weights. Anything else produces a cleaner table at the expense of accuracy.