Adobe gave its Firefly AI Assistant two new controls at MAX on Tuesday, and both are worth switching on before your team touches a client brief. One makes the assistant show you its plan before it spends anything. The other lets you decide which AI model does the work. Here is the catch: out of the box, both are set to let the assistant decide, and one of them resets every time you open a new session.
That matters more than it sounds, because in Firefly the model is no longer a style choice. It decides whether Adobe calls the output safe for commercial use or tells you the call is yours. It also decides what the job costs. If a social team running Black Friday variations never opens that menu, the assistant is making a legal and budget decision on their behalf, one prompt at a time.
This is a docs-based review. I read Adobe’s November 10 MAX announcement, the assistant’s help pages (all updated November 10), Adobe’s partner model and credit tables, and three independent write-ups, all fetched on November 12. Trying the free tier would have meant creating an Adobe account, so I did not, and nothing below is a claim about output quality.
What it is
Firefly AI Assistant is the chat layer on top of Adobe’s Firefly studio. You describe the job in plain language, it works out the steps and runs them using Firefly tools, capabilities borrowed from Photoshop, Lightroom, Illustrator and Premiere, and a menu of AI models from Adobe and its partners. Adobe says it now runs inside Firefly, Firefly Boards, Firefly Design and Firefly Graph. It is still labelled beta, as the assistant FAQ (updated November 10) says at the top of the page, and OpenTools’ November 10 access breakdown confirms that only four of the new Firefly tools shipped as generally available. The assistant is not one of them.
Access is broad on paper. The FAQ lists every paid Creative Cloud plan plus a subset of free accounts, adults only. It is available but “not optimized” for enterprise All Apps seats and higher education, and it does not exist in Adobe’s desktop or mobile apps, a handful of legacy enterprise plans, or K-12. If your designers live in desktop Photoshop all day, this is a browser tool they will have to go and open.
What changed on November 10
Three things, per the November release notes and the Next Web’s launch report the same day.
Plan mode. The modes help page (updated November 10) describes two agentic settings. Auto “lets the AI Assistant run a task on its own without asking for your approval.” Plan makes it lay out the steps and wait for you to review, edit and approve them, which Adobe says gives you visibility “before you consume credits or modify assets.” To get there you click the setting that reads Auto in the prompt field and pick Plan.
Model Selector. In the same panel there is a Customize toggle. Off, “the AI Assistant automatically chooses a model based on the task.” On, you tick only the image and video models you want it to use.
Preferences, Memory and brands. You can now save standing preferences, the assistant can learn from your chats, and it can pull a brand out of a guideline PDF, generate against it and check images against it.
What works
Plan mode is the right idea, done plainly. A multi-step job in this tool can mean retouching a batch of product shots, generating packaging mockups and cutting a short reel. Seeing the steps before anything runs is exactly the checkpoint a creative lead wants, and it is the first thing I would expect a careful producer to turn on.
The model picker is also real control, not decoration. You can lock the assistant to a short list in Settings, and Adobe’s own example is a team that wants Firefly Image 5 and nothing else. That gives brand teams a way to write a rule (“Firefly models only for anything that ships”) and have the tool enforce it rather than relying on everyone to remember.
Adobe is also unusually direct about the line between its models and everyone else’s. The FAQ says outputs from Firefly models inside the assistant “are considered safe for commercial use.” It says partner model outputs are your call. That is a clear sentence a legal team can work with, which is more than most AI vendors offer.
And the brand piece is useful for small teams. The brand guide (updated September 2) says the assistant saves logos, colour swatches, fonts and voice and photography guidelines to a Creative Cloud brand, then applies them automatically to later brand prompts. For a five person shop, not restating the guidelines every time is the whole appeal.
What breaks
The defaults point the other way. The FAQ is explicit: the assistant “automatically selects across both Firefly and partner models based on your request, unless a model is explicitly specified.” Agentic mode starts on Auto. So a new user who types a request and hits enter has let the assistant pick both the route and the model, and has skipped the plan.
Plan does not stick. The modes page says both the agentic setting and the model setting “persist only in the session you are currently in.” Model choices can be made permanent in Preferences. Plan cannot, as far as the docs say. Nadali’s November 11 read of the same pages reached the same conclusion and told readers to check the setting every time they start a session. That is a habit, not a control, and habits fail at 6 p.m. on a deadline.
The model decides your exposure. Adobe’s partner models page (updated November 10) lists what the assistant can reach today: Google’s Gemini 3 and Gemini 3.1 image models (the ones marketed as Nano Banana Pro and Nano Banana 2), Black Forest Labs’ FLUX.2, Google’s Veo 3.1 for video, Topaz Bloom for upscaling and ElevenLabs for speech. The same page says, in a note at the top, “You are responsible for determining whether a model is appropriate for your project.” Adobe does offer indemnity on some partner models, but its enterprise partner model page (updated October 23) limits that to generally available Google and OpenAI models inside Firefly Creative Production for Enterprise, covers certain copyright claims only, and says it should not be assumed to apply to other Adobe products. Nothing in those pages extends it to the assistant.
The bill shows up after the work. The FAQ says credit usage “varies by action” and that you check what an action used by hovering over the credits icon “after an action is completed.” It also says “Credit rates are introductory and may change.” Partner models carry their own posted rates. Nano Banana Pro is listed at 40 credits per generation, FLUX.2 at 20, and Veo 3.1 at 50 credits per second of video, all marked “for a limited time.” By my arithmetic an eight second Veo clip is 400 credits before a single retry. Plan mode shows the steps, but nothing in the docs says it shows a credit total before you approve, and Nadali makes the same point: approving a plan is not a spending limit. Topping up is not instant either. Adobe’s Zorana Gee told the Next Web that extra credits currently come only as a monthly add on, and one-off top ups are being evaluated with “no final answer.”
Memory is on, and it learns quietly. The Memory FAQ (updated November 10) says memory is activated by default, that the assistant learns what you like as you chat, and that “when it saves a memory on its own, it may not tell you each time.” One of the things it can remember is “whether you want to be asked before it creates something.” Read that next to the Plan setting and you can see the risk: a designer who says “just do it” a few times on Friday could be teaching the tool that this account does not want to be asked. Deleting a chat does not delete what it learned from that chat, and memories can be reworded as others change.
Brands apply themselves. The brand guide says the assistant “automatically applies brand assets created earlier” and you must name a different brand to switch. That is fine for one brand. For an agency login touching six clients, it means the wrong palette is one forgotten sentence away.
Who it is for
Small in-house brand and social teams on paid Creative Cloud plans get the most out of this right now: one or two brands, high volume of resizes and variations, and a creative lead who can set the rules once. Freelancers who already pay for Creative Cloud should try it on low stakes work.
Agencies can use it, but only with discipline about brands and memory per login. Enterprise teams on All Apps should wait for Adobe to say it is optimized for them, and anyone whose legal team needs indemnity on every shipped asset should treat partner models as off limits in this tool until Adobe says otherwise in writing.
Pricing and limits
The assistant comes with paid Creative Cloud plans and spends the same monthly credits as the rest of Firefly. Free accounts that have it get a limited number of daily generations that reset at midnight GMT. Adobe publishes no per-action price for the assistant itself, only that most generating and editing actions use credits at introductory rates. Partner model rates are posted on the partner models page and change by model, resolution and clip length. Custom skills are not supported yet. Always-on background agents are a preview, with beta promised for early 2027, per the Next Web. Admins can control access to the assistant, and enterprise Edition 4 and 5 admins can restrict partner models in the Admin Console.
What to do this week
Write the model rule, then set it in Preferences. One sentence: which models are allowed for work that ships, and which are fine for mood boards. Then have every seat save that list in AI Assistant Settings, Preferences, so it survives a new session. If you are on an enterprise edition that allows it, set it in the Admin Console instead of trusting each seat.
Make Plan the first click. Because it resets per session, put it in the brief template and the onboarding note. Anyone running a batch job starts by switching Auto to Plan and reading the steps, including which model each step names.
Audit memory on every shared login. Open Settings, read what the assistant has saved, and delete anything about skipping approval or anything that belongs to another client. If a login is shared across clients, turn Memory off for it.
Name the brand in every prompt on agency seats. It costs five words and removes the wrong palette risk.
Run a credit test before the holiday rush. Pick one realistic job, run it in Plan with your allowed models, and log the credits each step used from the hover counter. Multiply by your expected volume through Cyber Monday and compare it with what your plan includes, since a top up is a monthly add on for now.
Send your Adobe rep two written questions. Does any indemnity apply to partner model outputs made through the assistant on your plan? And will Plan mode ever show an estimated credit total before approval, or become a saved default? Get both answers in writing before the assistant goes anywhere near paid media.
The new controls are good. They are just not on until you turn them on, and one of them turns itself back off every session. Treat that as a setup job for this week, not a feature to admire.
