Kling 4.0: Runway API, Flash Trials and Video Costs

Runway lists model kling_4 for text, image and reference-guided generation, plus video editing. Its model catalog gives 720p and 1080p output with a 3–30-second duration. Use the model-specific endpoint controls rather than importing the full Kling application’s 4K, HDR or audio options into an API request.

Runway’s endpoint page checked October 26 advertises 20% off through November 30, 2026. Credits cost $0.01 each, before applicable tax. The table separates the offer from list prices. The page gives no cutoff timezone; use the current quote when submitting near the offer boundary.

The examples below use list prices. With input video, Runway bills output seconds plus the longer of input and output duration: O + max(I, O). At 720p, a ten-second output without video input is 157 credits ($1.57). With a five-second input it is 20 billed seconds × 9.4 = 188 credits ($1.88); a fifteen-second input makes it 25 × 9.4 = 235 credits ($2.35). Runway’s endpoint cost function rounds a fractional credit up to a whole credit.

Start from Runway’s model-specific example, use your own server-side API key and inspect the selected model and quoted credits before submitting. Record references, duration, resolution and all retries. For spoken dialogue, check whether the selected Runway endpoint supports native audio before submitting a job.

Kling AI’s September 28 announcement made Flash available to Ultra Yearly subscribers and said the full Kling 4.0 release was coming in October. The Flash creator tests below use that earlier access route.

T.J. Lovelady’s Flash first look describes the preview as 720p with generations up to 20 seconds. The full model adds the longer timeline and reference controls described below.

Kling’s September 30 comparison describes the full October release, not the Flash preview. Its main change is control over a longer sequence: place visual milestones along the timeline and combine references for the product, character and movement.

For a campaign, use keyframes for the reveal and final composition, then references for appearance and motion. For a three-second B-roll insert, those controls may add little. Both generations support 4K; the advertised upgrade also includes 10-bit HDR at 1080p/4K.

On September 29, Lynn Zeng posted a Flash text-to-video experiment focused on emotional dialogue delivery. She says the shown clip was the first generation with no rerolls. Watch the pause, facial expression and spoken delivery for the result she obtained in that scene.

Theoretically Media’s October 1 Flash review covers dialogue, an image-to-video diner scene and Omni Reference shots. Its chapter headings describe improved lip-sync drift relative to Kling 3.0, but also flat performances in one noir scene. The review includes an affiliate link.

The recurring practical test is whether the scene remains usable through movement and dialogue. Read the exact spoken line, track the character’s face during camera changes, and check the last seconds of the clip. Include a wide shot and a longer exchange if those are part of the finished scene.

Reddit user Jack_P_1337 liked Flash’s movement, instruction following and multilingual voices, but reported unwanted music despite explicit instructions, cuts during a requested continuous shot, and imperfect element consistency. The author also reported a start-frame input without the paired start/end-frame workflow.

In a separate thread, Isedo_m was disappointed after experimenting with keyframes and still preferred Seedance 2.5. Jack_P_1337 identified the version as Flash and suggested using Kling 3.0 for short connector clips between segments. Treat the connector workflow as an editing idea to try on your own sequence.

If your brief requires silence or one continuous take, test those requirements before subscribing for a longer project. Inspect the full clip for extra music and unrequested cuts, then confirm that the account supports the exact frame controls your edit needs.

Open Kling AI, confirm that your account has Flash access, and start with a short text-to-video scene. Specify who speaks, the exact line, the emotional change, the framing and the sound you want. Keep one difficult requirement per first trial so you can identify what caused a failure.

For example, try: “A static medium close-up in a quiet kitchen. One adult hears unexpected good news, pauses, then says ‘You really made it?’ with relief. Keep the camera still and the same face throughout. Room ambience, no music.” Save the prompt, chosen duration and credit charge with each result.

Then test the same brief with a first-frame reference. Compare the identity, delivery and continuity across several attempts. If you need production dialogue, budget for rerolls and the human review of names, numbers and pronunciation.

Lovelady’s pinned comment reports 120 credits for one 20-second 720p Flash video, which he converts to about $0.74 on his Ultra plan. The dollar estimate uses his Ultra subscription’s credit allowance.

At that reported credit charge, three attempts cost 360 credits and five cost 600. Count the attempts needed for an accepted scene, then add any editing, extension or export charge shown in your account. Check the account’s current credit schedule before committing to a larger batch.

Flash is a subscription-gated preview. FLUX 3 Video is available through BFL’s pay-as-you-go API, with explicit per-second rates and a draft-to-enhance workflow. Access and billing therefore differ before you compare image quality.

Give both models the same scene, dialogue, duration and reference assets. Compare accepted clips rather than the most attractive example from each feed. Track rerolls, generation time, identity drift and sound corrections. Match settings and record the selected model versions alongside the results.