Luma Agents

Luma Agents is the agentic creative workspace from Luma AI, a Redwood City, California company founded in 2021. Inside the browser-based Luma App and an iOS app, the agents plan, generate, iterate and refine creative work across video, image and audio from a brief, advancing several directions in parallel while keeping shared context across formats and team members. Luma names Publicis Groupe, Serviceplan, Mazda, Dentsu and Humain among the teams using it, and publishes case studies from Aperture, Serviceplan, OneDay, Uncharted and Boundless. The agents run on Luma's own models, Ray3.2 for video and Uni-1.1 for image generation and multimodal reasoning, and orchestrate third-party models inside the same workflow, including Veo 3.1, Kling 3.0 and Omni, Seedance, MiniMax, Flux, GPT Image, Nano Banana and ElevenLabs audio. Ray3.2 adds up to 16 keyframes per clip, video-to-video up to 20 seconds, reframing, Draft Mode and native HDR with 16-bit EXR export for grading and VFX. Newer tools include Luma Scenes, which lays a brief out as a storyboard so each key frame can be approved before rendering, plus Layers, Pages and reusable Skills. Developers reach the same models through the Luma API, with Python, TypeScript and Go SDKs, separate image and video endpoints, and a Build tier billed per use or a Scale tier with dedicated capacity, a latency SLA and a no-train guarantee. Luma's models are also offered through platforms including Envato, Comfy, Runware, Flora, Krea, Magnific, Fal and LovArt. There is no native CMS, DAM or social publishing integration, so finished assets are exported. No MCP server was published as of September 2026: the Luma API documentation and the main site carry none, and the MCP endpoint on Luma's older ReadMe documentation site is the documentation platform's search server, not a Luma capability. Pricing is in USD. Plus is $30 per month, or $25 per month billed yearly ($300), with 10,000 credits, third-party models, guest collaborators and commercial use. Pro is $90, or $75 yearly ($900), with 40,000 credits and four times the Luma Agents usage, and Ultra is $300, or $250 yearly ($3,000), with 150,000 credits and 15 times the usage. The pricing page lists Team, with team management, shared credits, usage analytics and SSO, and Enterprise, with commitments, training and custom fine-tuning, as contact-sales plans, although Luma's own llms.txt says the plans are Plus, Pro, Ultra and Enterprise with no Team plan. API use is billed in credits per model and output. What the entry tier does not buy: credits run out fast on premium video. A 5-second Ray3.2 clip costs 100 credits at 720p and 400 at 1080p, and HDR doubles or triples that, so the 10,000 monthly credits on Plus cover about 25 five-second 1080p clips. Third-party models are metered too, from 24 credits a second for Seedance 2.0 at 480p to 280 a second for Veo 3.1 with audio. Team management and SSO sit on the contact-sales plans. Luma Agents suits agencies, brand teams and production studios that want one workspace to take a campaign or film from brief to finished video and imagery. Teams that want articles turned into branded video should look at Lumen5, teams generating social video from a prompt should compare Invideo and Magic Hour, teams cutting long recordings into clips should consider Opus Clip, and teams that need social posting built in should look at Predis.ai. Current state Q3 2026: Luma launched the Uni-1.1 API on May 5, 2026 and Ray3.2 on June 9, 2026, then added Layers in July, Luma Scenes in August and Seedance 2.5 and MiniMax H3 as in-app models. In April 2026 the South African agency Boundless delivered Mazda's first AI-produced commercial with Luma Agents, Luma submitted 21 AI-generated finalists to Cannes Lions from 400 ads made in under eight weeks, and Luma and Wonder Project launched Innovative Dreams, a production services company backed by Amazon Web Services. Luma raised a $900 million Series C led by Humain in November 2025. On Plus, Pro and Ultra, Luma's terms allow it to train its models on customer inputs and outputs, while Enterprise content and the API's Scale tier are excluded from training.