GPT-6 Astra does the thinking and the driving; Higgsfield makes the images and the video. The pair works — but there are a few places in the middle that quietly cost you money if nobody tells you first. This is what to know before you start: the real costs, the six-step install, and what to do when it goes quiet.
Two paid subscriptions, and neither free tier is enough
Most guides skip this part, and it decides whether you can do any of it. Both sides cost money, and neither free tier opens this workflow.
- ChatGPT, for Astra: Picking GPT-6 Astra from the normal chat model list needs Pro, Business or Enterprise. On Plus ($20) you still get Astra, but only inside Work and Codex modes, not regular chat. Plus gives roughly 5 to 45 Astra messages per 5 hours, against builds that run 20 to 40 minutes — so treat every message as expensive, because it is.
- Higgsfield, for the visuals: A paid plan is required, in their words, and the free web tier does not open the plugin. Starter is $19 a month for 270 credits; Plus is $59 ($47 annual) for 1,200; Ultra is $129 ($99 annual) for 3,000.
Don’t save money on Starter. Starter has no access to Seedance 2.5 — the model behind their own flagship product and their UGC recipes. Pick Starter and half of this guide is closed to you. Plus is the real entry point.
Every price and limit below was accurate when this was written, in September 2026. Tools like these change their plans often — check the current ones yourself before you pay.
Three things that quietly burn credits
- "Unlimited" doesn’t apply through the plugin. Their own wording: every generation through MCP, the ChatGPT plugin or the CLI deducts credits at standard rates. Unlimited applies on their website only.
- Every published price is a first attempt. Their own blog puts the real figure at three to ten attempts per usable clip. A generation that finishes but comes out wrong is not refunded, because technically it succeeded. Only generations that return nothing get credited back.
- There is no hard cap. Nothing stops when you have spent enough, so the guard has to come from you.
In money: on Plus, 1,200 credits costs $59, so a credit is about 5 cents. A 15-second UGC ad at 1080p is 270 credits — roughly $13, and 22% of your month. A 30-second cinematic single take is 540 credits, roughly $27, or 45% of your month. Those are their own published figures from their own runs — multiply by five for what you will actually spend.
From now on, before you generate any image or video, tell me roughly how many credits it will cost and why, then stop and wait for me to reply "OK". Generate nothing I haven't confirmed.
Before you generate, write out what you're about to make, in words: - the shot, and what's in frame - length, aspect ratio, resolution - which model, and why that one - the estimated credit cost I may want to change it. Generate only once we've settled it — don't iterate by burning generations.
Installing the plugin, in six steps
- Open the plugin directory: chatgpt.com/plugins, or Plugins in the left sidebar.
- Search for Higgsfield.
- Add the plugin.
- Sign in to Higgsfield.
- On ChatGPT Plus, switch to Work mode — that is where Astra lives for you. Then type @ and pick Higgsfield to pull the plugin in.
- For anything visual, set reasoning to Extra High. For pure code builds, Medium is usually the better trade — High mostly just makes you wait.
I want to use Higgsfield to make [what you're making]. Before we start: 1. Confirm you can actually reach the Higgsfield plugin right now 2. Tell me which model you plan to use, and why that one 3. Quote the credit cost before each generation and wait for my OK Tell me when you're ready. Don't generate anything before we've agreed the plan.
Connected, but nothing happens?
- It connects and stays silent: Start a brand new conversation and name Higgsfield explicitly in your first message.
- You see Authorization with the MCP server failed: Go to Settings, then Plugins, then Higgsfield, and reconnect.
- It generated but you can’t find it: Results land in your Assets in about 60 seconds.
Try it: a UGC clip that doesn’t read as AI
Most AI video gives itself away by being too clean — too smooth, too centred, too lit. This prompt works the other way: the phone sits crooked and nobody straightens it, the framing is slightly off, the typing rhythm is irregular, there are pauses, and she never looks at the camera. The imperfections only happen if you ask for them by name.
Replicate the person in the reference image exactly: face, skin detail, hairstyle and clothing, strictly as shown, adding nothing and changing nothing. Carry over the person's identity only — not the lighting. Keep the imperfections on purpose. The propped-up phone sits a few degrees crooked and she never straightens it. The framing is slightly off-centre. Keep the natural room noise: no music, no studio-style polish. She sits at a computer working, body relaxed, focused on the screen, both hands typing naturally — the rhythm irregular, pausing now and then to think. A few seconds in, one hand leaves the keyboard and reaches naturally for the coffee cup beside her, her eyes still on the screen. She takes a small sip, presses her lips together slightly, sets the cup down, and puts her hand back on the keyboard to keep typing. She never looks at the camera, never smiles for it, never poses. Keep the small movements of someone actually working: blinking, light breathing, fingers pausing, small shifts of gaze, natural changes in posture. The motion should not be too smooth or too perfect — it should look like a real clip someone caught on a phone.
Before you use it: the reference photo should be you, or someone who has agreed to it.
Then try: swapping the background behind a talking head
No moving house, no tidying the desk, no studio rental. Shoot a talking head, then put it in a modern tech-company studio — the person untouched, only the room changed. Most of this prompt is spent on what not to change: face, hair, clothes, movement, mouth shapes while speaking, camera angle, timeline. The way background swaps go wrong is the model helpfully regenerating the person too.
Replace the background of the talking-head video I'm providing with a modern, professional, high-end tech company office. The person must stay exactly as they are — face, hair, clothing, build, movement, mouth shapes while speaking, gestures, camera angle and the original timeline. Do not regenerate or alter the person. The new background is a cinematic modern creative studio: floor-to-ceiling windows, soft natural daylight, a clean desk, a computer screen and a few subtle tech details, with natural depth of field and light bokeh behind. It should feel like a professional AI or tech company's creator studio — premium, real and uncluttered, not like a stock photo. Re-match the lighting, shadows, colour temperature and edges between the person and the new environment so they sit in the scene naturally. Keep every movement of the person and all camera motion from the original. Change the background only.
Don’t cut the last paragraph — the one about re-matching light, shadow, colour temperature and edges. Without it the person looks pasted on, and everyone can tell.
Two things that don’t work inside ChatGPT
Standalone audio generation and Website Building do not work through the ChatGPT plugin. You get spoken dialogue baked into a video generation, but no separate voiceover, voice cloning or dubbing. For those, use the Higgsfield website, or connect it to Claude instead.
If you remember one thing: Multiply every published price by five before deciding a clip is worth making. The first-attempt price looks great, and you will not get it right on the first attempt.
