On October 1, Adobe changed two things about Firefly AI Assistant at once. It opened the beta to anyone with a Creative Cloud plan, and it started charging those people’s plan credits for the work the assistant does.
Either change alone would be routine. Together they turn a free sandbox into a production tool with a running meter, and they land in front of a much wider set of seats than the people who tested it this summer. The part that matters most for marketing teams sits in one line of Adobe’s FAQ: the assistant picks the model for you unless you tell it otherwise.
A note on evidence before going further. This is a primary-source review. I have not run the assistant on a paid account for this piece. Everything below comes from Adobe’s help center, product page, and launch blog, cross-checked against reporting from TechCrunch, The Verge, and CreatorStack. Where Adobe’s own pages disagree, I say so.
What it is
Firefly AI Assistant is a conversational agent inside the Firefly web app. You describe an outcome (“turn this product shot into a set of social crops,” “retouch these headshots,” “put this logo on this box”), and the assistant plans the steps, picks tools from across Adobe’s catalog, and runs them. Adobe says it draws on capabilities from Photoshop, Lightroom, Illustrator, Premiere, and Firefly itself, and the April launch post talked about more than 60 pro-grade tools arriving as the beta rolls out.
It grew out of Project Moonlight, which Adobe previewed at MAX last year, according to TechCrunch. It was announced on April 15 and went into public beta on April 27 for Creative Cloud Pro and paid Firefly plans.
The pitch is that you no longer need to know which tool in Adobe’s large catalog does the job. You describe the result, and the agent finds the tool, which is useful for a social manager who has never opened Lightroom and for a designer who would rather not do the twelfth resize by hand.
What changed on October 1
During the spring and summer beta, Adobe’s launch post promised eligible users complimentary daily generative credits “exclusively for use with the assistant,” refreshing each day. That was the sandbox.
Adobe’s FAQ, last updated September 24, now says two things kick in from October 1. First, the assistant becomes available to all users with a Creative Cloud plan, not only Creative Cloud Pro and paid Firefly tiers. Second, for paid plans, “certain AI Assistant actions will use your available AI credits.” Usage varies by action, and the only way to see what an action cost is to hover over the credits icon after it finishes. Adobe calls the rates introductory and says they may change. CreatorStack, reviewing the same documents on October 4, reached the same reading: broader access, credit use on certain actions, and no universal price per project.
So the beta label stays, but for paid users the FAQ no longer describes a separate allowance. The cost of certain actions now comes out of the same monthly pool your team already uses for video, premium features, and partner models.
What works
The workflow idea is sound, and Adobe’s documentation suggests it was built for real production tasks rather than demos.
Skills are the strongest piece. Adobe ships prebuilt workflows such as Portrait Retouch, Social Media Assets, and Mockup Studio, which you can call explicitly with a slash command or let the assistant trigger when it recognizes the intent. You can also tell it to skip skills and keep things simple. For repetitive marketing output (one hero asset reframed for every placement, a batch of event photos brought to a consistent look) this is exactly the work that eats junior designer hours.
You stay in the loop. Adobe’s product page and launch post both stress that you see the steps the assistant takes and can redirect at any point. Outputs save to Creative Cloud storage, so a designer can open the result in Photoshop and finish it properly.
The data posture is clear. Adobe says your content is not used to train its models or partner models, and that the assistant only stores preferences you explicitly ask it to remember.
Failure handling is fair. Credits for a failed generation are typically refunded within minutes, and up to 24 hours for long video runs, per the credits FAQ.
Where it breaks or is fenced
This is the longer list, and it is where operators should spend their attention.
The router decides the model, and the model decides almost everything else. Adobe’s FAQ says the assistant “automatically selects across both Firefly and partner models based on your request, unless a model is explicitly specified.” Adobe’s partner models page lists the assistant as a surface for Google’s Gemini image models (Nano Banana Pro and Nano Banana 2), Black Forest Labs’ FLUX.2, Google’s Veo for video, Topaz Bloom for upscaling, and ElevenLabs for speech.
That routing choice carries three consequences most people typing a prompt will never think about:
- Cost. Partner models are premium features that spend credits at their own rates.
- Commercial safety. Adobe says outputs from Firefly models are safe for commercial use. For partner models, “you are responsible for determining whether a partner model is appropriate for your project.”
- Indemnity. The credits FAQ says that for eligible teams and enterprise customers, beta outputs are not indemnified. The assistant is a beta.
Put plainly, a coordinator asking for “a lifestyle version of this product shot” may get a partner-model image, spend partner-model credits, and produce something your agency contract assumes was covered when it was not. Nobody made a bad decision. The router made it for them.
Rates are visible only after you spend them. There is no published per-action price list for the assistant. You learn the cost by hovering after the fact. Adobe’s standalone rate card gives a sense of scale for the models the assistant can call: Nano Banana Pro at 40 credits per image and Veo 3.1 at 50 credits per second of video, both marked limited-time, and ElevenLabs speech at 15 credits per 1,000 characters. Those are Firefly rate card numbers, not confirmed assistant charges, and Adobe says assistant rates are introductory. Still, the arithmetic is useful. At the rate card price, one eight-second Veo 3.1 clip is 400 credits. A Creative Cloud Pro seat gets 4,000 premium credits a month.
Commercially safe does not mean cheap. On Adobe’s own credit table, Firefly’s video model lists at 100 credits per second at 1080p, double the current Veo 3.1 rate. Pinning to Firefly models is a rights decision, not a savings plan.
Wider access meets thin wallets. Opening the door to “all users with a Creative Cloud plan” sounds generous until you read the credit table. Creative Cloud Pro and the Firefly plans include unlimited standard generations plus a premium pool (2,000 to 50,000 a month on Firefly tiers, 4,000 on Creative Cloud Pro). Creative Cloud Standard includes 25 credits a month. Single app plans bought on or after June 17, 2025 also get 25. A team where half the seats are single app licenses will find out quickly that the assistant is available to everyone but affordable for a few.
Credits do not pool or roll over. Unused credits vanish at the monthly reset, and on teams and enterprise plans they are not shared across users. Only certain enterprise agreements can buy a shared pool. The heavy video user and the light user each live on their own allowance.
Surface and plan gaps. The assistant runs in Firefly on the web only, not in Adobe’s desktop or mobile apps. Several legacy enterprise plans are excluded outright, and Adobe says it is available but not optimized for enterprise all-apps subscribers and for languages beyond English, French, German, and Japanese.
Promised features are still pending. At the April announcement, The Verge reported that creatives would be able to create their own Creative Skills. Adobe’s current FAQ says custom skills are not supported yet. Preferences live at the user level with no screen to view them, so there is no way to set a brand or client rule at the project level today.
Adobe’s own pages disagree. The help FAQ says every Creative Cloud plan qualifies from October 1. The product page still says the beta is only for Creative Cloud Pro and Firefly Pro, Pro Plus, and Premium. Check the actual account before you plan around either.
Who it is for right now
It fits best for in-house social and content teams on Creative Cloud Pro or Firefly Pro and above, doing high-volume adaptation work: reframes, resizes, batch retouching, mockups for internal review. That is where the credit pool is deep, the tasks are repeatable, and the rights stakes on any single output are modest.
It fits less well, for now, for agencies delivering final paid media to clients under indemnity terms, for enterprise teams whose admins have not set partner model access, and for anyone on a 25-credit plan who expects to make video.
Enterprise customers on Edition 4 and Edition 5 have one real lever. Adobe says those plans can control which partner models are available to which users from the Admin Console. That is the cleanest way to stop the router from choosing something your legal team has not approved.
What it costs
There is no separate price for the assistant. It spends the generative credits already in your plan. The monthly allotments Adobe publishes are 4,000 premium credits on Creative Cloud Pro, 25 on Creative Cloud Standard, 2,000 on Firefly Standard, 4,000 on Firefly Pro, 10,000 on Firefly Pro Plus, and 50,000 on Firefly Premium. When paid users run out, Adobe’s options are to upgrade, buy more credits, or wait for the reset. Free accounts in some regions get a small daily allowance that refreshes at midnight GMT.
What you cannot do yet is forecast a project. TechCrunch noted in April that Adobe had not said how the assistant would be priced. Six months later the answer is “per action, introductory, shown afterward,” which is a meter without a published price list.
What to do this week
The useful move is to treat October 1 as a policy change, not a feature release, and handle it the way you would any new spend line.
- Find out who has it. Ask your Adobe admin which seats now see AI Assistant in Firefly on the web. If the help FAQ is right, that is every Creative Cloud seat.
- Write one model rule into every creative brief. For anything client-facing or paid, the brief names the model (“Firefly models only”) unless legal has cleared a specific partner model. Adobe says the assistant respects an explicit model choice, so make naming it a habit.
- Run a metered pilot on one real job. Pick a single repeatable task, such as one hero asset reframed for every placement. Run it end to end, hover the credits icon after every step, and log model, action, and credits. Do the same job once with a model pinned and once without, and compare.
- Set the partner model controls if you can. Enterprise Edition 4 and 5 admins should decide this week which partner models are allowed and for whom, before usage habits form.
- Keep beta outputs out of indemnified deliverables. Until the assistant leaves beta, agencies should treat its output as draft material that a human finishes in a non-beta tool before it ships to a client.
- Match seats to work. Put heavy video and partner model work on the seats with deep pools, and stop light seats from discovering a 25-credit ceiling halfway through a campaign.
Adobe built an agent aimed at real production work and then, quietly, handed procurement and rights decisions to whoever is typing. The fix costs nothing but a sentence in the brief and an afternoon of logging. Do it before the end-of-month reset teaches the lesson for you.
