The cost math — when AI video is viable and when it bankrupts you — step 4 of 6

A 60-second promo, priced end-to-end across every model

Concrete scenario: a 60-second promotional video for a SaaS product. Twelve shots of 5 seconds each. Image-to-video pipeline (frames composed with gemini-3.1-flash-image at ~$0.067/frame). Usable B-roll quality bar — not Super Bowl, but client-presentable. Retake rate of 2.5x is realistic for I2V at this quality bar.

The math, line by line

Frame composition spend (constant across models): 12 frames × $0.067 ≈ $0.80

Video generation spend (varies by model): 60 seconds × price_per_sec × 2.5 retake multiplier

Sora 2 jobs are 16-20 seconds, so a minute is three extends. The dollar column is still raw seconds × rate. The $30 at 1024p is 60 × $0.50, not one request.

The full table

Model	$/sec	Video gen × 2.5	+ frames	Total
Veo 3.1 Lite (720p+audio)	$0.05	$7.50	$0.80	$8.30
Runway gen4_turbo I2V	$0.05	$7.50	$0.80	$8.30
Kling 3.0 standard	$0.084	$12.60	$0.80	$13.40
Omni 1.1 Flash (720p)	~$0.10	$15.00	$0.80	$15.80
Veo 3.1 Fast (720p+audio)	$0.10	$15.00	$0.80	$15.80
Sora 2 (720p)	$0.10	$15.00	$0.80	$15.80
Runway gen4.5	$0.12	$18.00	$0.80	$18.80
Kling 3.0 Pro	$0.168	$25.20	$0.80	$26.00
Sora 2 Pro (720p)	$0.30	$45.00	$0.80	$45.80
Veo 3.1 Standard (audio)	$0.40	$60.00	$0.80	$60.80
Sora 2 Pro (1024p)	$0.50	$75.00	$0.80	$75.80

At retake 1.0 the Lite row is $3 of video + $0.80 of frames. That is the only way the $3-5/finished-minute headline holds.

What this table tells you

Three observations.

The cheap-to-premium spread is ~9x. $8 for Lite at 2.5x all the way up to $76 for Sora 2 Pro 1024p. Same brief. Sora 2 / Videos API is removed 2026-09-24.

There is no single "right" model. The answer depends on:

• What the client expects (cheap-tier output is recognizable to trained eyes and disqualifies premium client work).
• What the audience expects (TikTok audiences accept Lite output; brand marketers do not).
• Whether the piece needs audio (if yes, Veo Lite at 2.5x is $7.50 and Standard is $60 — same audio pass, different look. Omni also generates an audio track by default at ~$0.10/sec).

The retake rate dominates the choice. If you're disciplined about prompts (lesson 02) and have a real shot list, you might hit 2.0x retake instead of 2.5x. That's a 20% cost cut across the board. Conversely, if you're sloppy and hit 5.0x retake, every number doubles. The stills barely move the total: $0.80 against $7.50-$75 of video. Stack them for control.

What goes where in real production

Three rules from working agencies as of 2026-09-01:

1. Hero shot? Veo 3.1 Standard, or Omni if you want the default Google call. The output quality is non-negotiable for client work. ~$24/min raw on Standard, ~$6/min on Omni. Used for the one or two shots that are the deliverable. Sora 2 Pro is not the production default.
2. B-roll / supporting shots? Veo 3.1 Lite or Veo 3.1 Fast. Quality is good enough to cut next to a hero shot. $3/min raw on Lite, $6/min on Fast. Used for the supporting 70-80% of a piece.
3. Performance marketing tests? Veo 3.1 Lite or Runway gen4_turbo I2V. Quality is "AI-flavored" but the testing volume is the point. $3/min raw. Used when you need 50+ variants.

A typical 60-second promo for a serious client lands at $15-45 in real spend by mixing hero shots from Standard with B-roll from Lite. Pure all-Standard at 2.5x is $60+. Pure cheap-tier is ~$8 but won't ship as a hero.

The Higgsfield wrinkle, revisited

If the brief requires deterministic camera moves (the brand book specifies "all product shots are slow orbits," etc.), Higgsfield becomes the camera layer regardless of tier. If you don't need the camera control, the subscription is dead weight and you should go direct to generate_videos or interactions.create.

The script

Step 5 has you write cost_per_minute(model, retake_rate) — a function that takes a model name and a retake multiplier and returns dollars per finished minute. You'll use it to make viable/not-viable decisions on your own briefs without opening a spreadsheet.