Uni-1.1 API (correct name; production version of March's Uni-1) is now generally available — IMAGE only, not video. Two SKUs (Uni-1.1, Uni-1.1 Max), two endpoints (Generate / Modify), Python + JS SDKs, up to 9 reference inputs, ~31 s per image at ~$0.09 (2K). The '#1 ELO' claim is Luma's internal eval, not LMArena — and Luma is #2 in pure text-to-image.
Luma Labs has opened the Uni-1.1 API for general production use — a refinement of the Uni-1 model unveiled in March 2026 and now the company's flagship image-generation API. The release ships in two SKUs (Uni-1.1 and Uni-1.1 Max), two REST endpoints (Generate Image and Modify Image), and official Python and JavaScript SDKs. It accepts up to nine reference images per call to preserve identity, composition, or style.
Key facts
- Correct product name is Uni-1.1 API (the production release; Uni-1 launched March 5, 2026).
- Image-only at launch: PNG/JPEG, aspect ratios 1:1, 9:16, 16:9, ~31 s per image. Audio and video output are on the roadmap. Luma's video stack remains Ray3.14.
- Two tiers: Build (usage-based) and Scale (production rate limits, dedicated support).
- Pricing: ~$0.09 per 2K text-to-image, ~$0.0933 for editing / single-reference, ~$0.11 for 8-image multi-reference. 10–30% cheaper than Nano Banana 2 ($0.101) and Nano Banana Pro ($0.134).
- Architecture: decoder-only autoregressive transformer interleaving text and image tokens — distinct from diffusion-based competitors.
- "Uni" stands for "Unified Intelligence", a new family alongside Luma's existing Ray (video) and Photon (image) lines.
The "#1 ELO" claim deserves a footnote
Luma's marketing leads with a Human Preference ELO win — true, but with three caveats most coverage skips.
The "#1" is from Luma's own internal pairwise human-preference evaluation, not LMArena, Artificial Analysis, or any third-party board.
Where Luma actually claims first: Overall, Style & Editing, and Reference-Based Generation. Where Luma is honest about being #2: pure text-to-image — behind Google's Nano Banana family. Independent benchmarks confirm a real architectural advantage on reasoning: on RISEBench, Uni-1 scores 0.51 overall (narrowly tops Nano Banana 2 and GPT Image 1.5), 0.58 on spatial reasoning vs. 0.47 for Nano Banana 2, and 0.32 on logical reasoning vs. 0.15 for GPT Image 1.5 (The Decoder).
That's a defensible claim. The unqualified "#1 in human preference" framing is not.
Where Uni-1.1 sits in the market
The competitor set is image, not video. Uni-1.1 is positioned against Google Nano Banana 2 / Nano Banana Pro and OpenAI's GPT Image 1.5 — not Runway Gen-4, Pika 2.x, Veo 3.5, or Sora 2 (those are video models). The architectural differentiator is decoder-only autoregressive generation interleaving text and image tokens, which is what produces the measurable reasoning lift on RISEBench.
Production traction
The more defensible signal is who's already wired in. Uni-1.1 ships with named integrations across Envato, Krea, Fal, Magnific (formerly Freepik), Runware, Comfy, Flora, and LovArt. Earlier Uni-1 enterprise customers include Publicis Groupe, Serviceplan, Adidas, Mazda, and HUMAIN. Luma's case-study claim is a $15 M ad campaign localised into 40 markets in 40 hours for under $20 K — the kind of throughput number that defines whether "production-grade" is a marketing word or a measurable one.
What to watch
- Whether Luma publishes the Human Preference ELO methodology and pairwise dataset so the #1 claim can be independently reproduced.
- When the audio and video roadmap materialises — bringing Uni into territory currently owned by Ray3.14, Veo, and Sora.
- Pricing pressure: a 10–30% undercut of Nano Banana is meaningful at agency throughput, and could force Google to respond on per-image cost.
Sources: Luma — Uni-1.1 API official announcement · Luma — Uni-1 product page · VentureBeat · The Decoder — RISEBench scores · MarkTechPost — autoregressive transformer detail · Artificial Analysis — independent video board (for context) · Winbuzzer
