By Gen Feeds Editorial
Adobe's MAX 2026 conference runs through Nov 12, with livestream blocks for the opening keynote, Sneaks and the Inspiration Keynote on its schedule[1]. The item most worth weighing sits below the stage, in a menu. Firefly, Adobe's generative studio, sorts its model picker into Adobe models carrying a "Commercially safe" marking, a "Partner models" tier, and a third tier labeled "Models created by others"[2]. The partner tier names Gemini Nano Banana models, GPT Image, Runway, Luma AI, Kling AI and ElevenLabs — tools from other vendors, listed in the same studio whose own models carry the safe marking[3]. One prompt bar holding Adobe's safe-marked models and its rivals' models together, with a documented way to compare them, is the pleasantly surprising part. The same pages that document the openness also draw its limits, and the limits are where the judgment lands.
The model picker now lists the competition
The partner-models page states the pitch without hedging: "With Firefly as your multi-model creative AI generator, you can design, edit, and experiment with top generative AI models, all in one place. Access Gemini Nano Banana models, GPT Image, Runway, Luma AI, Kling AI, and more."[3] The audio lineup names ElevenLabs in the same sentence as the image and video models[3]. The comparison mechanic is documented, too: "Run the same prompt across multiple AI models in Firefly to compare results side by side and choose the best fit for your project."[3]
The framing matches the menu. The Firefly page leads with "Generate and edit images, video, and audio all in one AI-powered creative studio"[2], and the partner page calls the same arrangement "the next evolution of creative AI... with image, video, audio, and vector tools"[3]. In the picker layout, those partner names sit one menu click away from Adobe's own models[2].
| Tier in the generation bar | What Adobe's pages say about it |
|---|---|
| Adobe models | Marked "Commercially safe"; Adobe says its own models cover images, video, audio and vectors[2] |
| Partner models | Named roster including Gemini Nano Banana models, GPT Image, Runway, Luma AI, Kling AI and ElevenLabs[3] |
| Models created by others | Listed as the third option in the picker[2] |
How deep the roster goes
The openness reaches well past the image picker, and Adobe's documentation treats it as infrastructure rather than an experiment. The help center maintains separate pages for generating images using FLUX, Gemini, GPT Image, Imagen, Ideogram and Runway[4], plus a page for generating vectors using partner models[4]. The audio stack is documented in the same pattern: sound effects from text prompts or a voice guide, music generation, speech from text, speech using partner models, and audio and video translations in over 20 languages[2][4]. Video documentation covers text-to-video, image-to-video, partner models, transparent backgrounds, translation, text-prompt editing and upscaling, with named partner pages for Kling and Seedance[4]. Adobe even attaches suggested roles to specific models: "Try GPT Image for photorealistic product shots, Kling AI for synchronized multishot videos, or Gemini Nano Banana models to refine prompts before generating final images."[3] The practical meaning is that model choice becomes a per-task setting inside one studio, across images, video and audio. Prompt length in the web generator caps at 750 characters[2], a reminder that the aggregation happens inside one controlled interface.
Firefly AI Assistant stays beta
Choosing a model per task is one half of the arrangement; the other half is the assistant layer that sits above it, and Adobe ships it explicitly unfinished. Adobe describes Firefly AI Assistant (beta) as a system that "will use tools from apps like Photoshop, Illustrator, and more to complete multistep projects in moments"[5], with a defined mechanic: "the AI Assistant creates a full execution plan before anything runs. Review it, adjust it to match your vision, and approve."[5] The tool span is named: Photoshop, Illustrator, Lightroom, Premiere, Acrobat and other Creative Cloud apps[5]. Adobe lists brand context, plan mode and model selector as new beta capabilities[5], ships prebuilt Creative Skills workflows such as Portrait Retouch, Social Media Assets and Mockup Studio[5], and has placed the assistant inside Firefly Boards[5]. The page also separates it from per-app assistants, which it says stay focused on tasks within a single app[5].
One line on that page sets the limit: "Video is a preview of upcoming technology. Not all features shown are currently available."[5] Adobe is telling readers its own demonstration footage outruns the shipping feature set. The scope is therefore narrow: plan-then-approve is the mechanic Adobe documents as working, and everything demonstrated beyond it stays in beta until Adobe says otherwise. The same page calls the results "production-ready"[5], which is Adobe's marketing claim sitting next to its own beta label, and the beta label is the one with operational force.
Where the commercially safe label stops
The trust boundary is visible in the picker layout. The "Commercially safe" marking attaches to the Adobe models tier[2], and Adobe's phrasing keeps the two tiers apart: "Adobe Firefly includes Adobe's own commercially safe generative AI models for images, video, audio, and vectors, plus partner models from Google"[2]. Partner models enter the studio without that marking. What the marking covers in licensing terms is spelled out, if anywhere, in the terms documents the product pages link to; the picker itself shows only the label, and the label travels with Adobe's own models.
Usage is a separate story. Adobe documents a generative credits system with a dedicated "Generative credits overview" page in its Firefly help center[6], alongside a marketing offer of "free daily generations for images, video, and audio in Firefly"[2]. The free headline and the credits system coexist on Adobe's own pages, and the pages cited here state no billing rate for partner-model runs. For the repeated-prompt comparison the partner page is selling, the meter is the open variable.
The surprise holds on strategy
Across the pages cited here, the arrangement is documented fact: rival models run inside Firefly's prompt bar with a side-by-side comparison mechanic, the safe marking stays on Adobe's own tier, and the assistant ships in beta with plan-then-approve as its documented mechanic. The pages document the arrangement, not its history. Read as strategy, Adobe is selling three things at once: its own models on the safe marking, rival models on one-stop convenience, and the beta assistant as the layer above both. That reading covers strategy only; model quality and licensing terms are outcome questions the cited pages do not settle. The pleasantly surprising part is where the openness lives: in the model picker, where Adobe's safe-marked models and its rivals share one list. If Adobe's own models keep the safe marking to themselves and repeated-prompt comparison stays workable on the credit meter, the open studio is the more interesting product; if either changes, the arrangement has a catch after all.