Point of view · August 2026
The model landscape, and why routing beats picking
There is no single best model — there are jobs, and models that are currently best at them, at a price. A production studio's edge is a routing table it re-evaluates monthly, not a favorite tool. This page is the live routing table behind the Studio demo.
Wired into this demo
| Model | List price | Job it owns here |
|---|---|---|
| Nano Banana Pro (Gemini 3 Pro Image) | $0.134/image | Promo tiles with live text, bilingual versioning, brand-critical hero shots. |
| Nano Banana (Gemini Flash Image) | $0.039/image | Format adaptations, seasonal variants, bulk versioning at scale. |
| Flux 2 Pro (Black Forest Labs) | $0.05/image | Photorealistic hero stills where the image is the product. |
| Flux Kontext Max (BFL, via fal.ai) | $0.08/image | Packshot challenger: identity-true edits where texture matters more than dense label text. |
| Seedream 4.0 Edit (ByteDance, via fal.ai) | $0.03/image | Packshot challenger: the value benchmark every bake-off should include. |
| Veo 3.1 Fast (Google) | $0.15/second | Drafting, social clips, high-volume video versioning. |
| Veo 3.1 (Google) | $0.75/second | Hero spots and finished broadcast-style deliverables. |
| Kling 3.0 (Kuaishou, via fal.ai) | $0.1/second | Cost-efficient social cutdowns; same product across many shots. |
| Runway Gen-4 (via fal.ai) | $0.15/second | VFX-leaning brand films and shot-heavy storytelling where you expect to iterate on the motion. |
| Seedance 2.5 (ByteDance, via fal.ai) | $0.46/second | Short-form product ads at volume; longer cuts no other model does in one pass. |
| Seedance 2.5 Reference (ByteDance, via fal.ai) | $0.46/second | Product-identity-critical ads: the packaging must stay exactly itself while the camera moves. |
List prices as of August 2026; they inform the Studio's pre-flight cost estimator and are kept in one config file so drift is a one-line fix.
Stills — the read
Nano Banana Pro (Gemini 3 Pro Image)
The reasoning-driven image model. Where it wins: text rendered inside images (packaging, promo badges, bilingual headlines), brand consistency across a series, and instruction-based editing. This makes it the backbone of retail versioning — EN/FR promo tiles are its signature use case here.
Nano Banana 2 / Flash Image
The volume tier. Roughly a quarter of Pro's price, fast, and good enough for format adaptations and seasonal rethemes where the hero already set the look. High-volume pipelines live or die on knowing when 'good enough' is good enough.
Flux 2 Pro (Black Forest Labs)
Best pure photorealism in market — food photography, texture, natural light. When the image itself is the product (hero shots, editorial), Flux earns its place. Weaker at in-image text, which is why routing matters instead of picking one 'best' model.
Ideogram (not wired in, on the radar)
The typography specialist. If the studio's flyer/promo tile volume grows, a dedicated text-in-image model is the next evaluation — a good example of the continuous pilot-and-adopt loop.
Video — the read
Veo 3.1 (Google)
The quality ceiling: native 4K, native 48kHz synchronized audio, best lip-sync available. The Fast tier at ~$0.15/s changes the economics of drafting — iterate on Fast, finish on Standard. Deep Gemini API integration means one vendor surface for reasoning + stills + motion.
Kling 3.0 (Kuaishou)
The value play at ~$0.10/s — 4-7x cheaper than premium tiers — with uniquely strong subject consistency across shots. That consistency is exactly what product content needs: the same can, jar or bag recognizable in every cut.
Seedance 2.5 (ByteDance) — reference-to-video
The answer to product drift, and the most interesting endpoint in the roster. Instead of one grounding frame, it accepts up to 50 multimodal references addressed positionally in the prompt as [Image1], [Image2]… — so you assign one reference the product's identity, another the palette, another the camera move. The packaging stays itself while the camera moves, which first-frame conditioning cannot guarantee. References aren't limited to stills either — a short clip can carry the camera move or the cut rhythm, addressed as [Video1] in the same prompt. It also renders native 30-second takes in a single pass, and generates audio jointly with the picture rather than after it.
Runway Gen-4
The creative-control surface: motion brush, camera control, video-to-video. Built around iteration rather than one-shot generation, which is why it stays the tool a studio team works in daily — also wired in here so the same brief can be pushed through it.
Sora 2 (OpenAI)
A cautionary tale, on purpose: OpenAI discontinued the app in April 2026 and sunsets the API in September 2026. Building on it would have stranded a pipeline. Vendor-risk assessment is a core studio discipline, not paranoia.
Ready / emerging / not yet viable
Ready to use
- Stills generation & editing for adaptations and versioning
- Text-in-image promo tiles (with a QA gate)
- Short-form video (5–8s) with native audio
- EN/FR headline localization with human review
Emerging — pilot now
- Multi-shot narrative spots (Seedance-class models)
- Automated brand-compliance checking
- Video restyling of existing footage (WAN-class)
Not yet viable
- Unreviewed publish-direct pipelines
- Long-form (30s+) fully generated spots at broadcast QC
- Photoreal likenesses of real people (policy, not tech)