Four images cost 7,000 credits. One costs 4,000
A Japanese walkthrough of Meta's Muse agent driving the PixAI API logged 4,000 credits for one image and 7,000 for four. The reading that matters is which four: the near-flat increment stops at one prompt.
AI-assisted draft. Reviewed and edited by the Phosphene team before publication.
On this page

Ai-Hakase, a Japanese creator who documents AI art workflows on note and YouTube, published a step-by-step walkthrough of connecting Meta's Muse agent to the PixAI API: get a key, put it somewhere safe, ask the agent to write the script, generate. The screens are dated 15 November 2026, and the whole thing is aimed at people who will never write code. You paste, you approve a permission dialog, you watch an image appear.
The useful part is not the setup. It is a number buried in the middle of the post.
The author generated four images in one task and 7,000 credits left the account. A single-image task on the same model and settings cost 4,000. The two configurations differ only in batch size, and the gap between them is 3,000 credits.
Then comes the detail that decides whether that gap is a real discount or a story about one slider. PixAI's API takes a single prompt per task, and batchSize is either 1 or 4. Those four images were four variations of one prompt. They were not four different prompts.
That distinction is the whole article. A batch buys more attempts at one decision for barely more money. It does not buy a cheaper way to produce four different pictures.
Where the increment stops
PixAI has not published the pricing formula, so nobody outside the company can say how it allocates the charge between the first image and the rest. What the two numbers do show is the task-level difference: configuring one task for four images cost 7,000, configuring it for one cost 4,000, which is seventy-five per cent more money for three hundred per cent more images.
The increment does not travel. A task takes one prompt, so twenty distinct prompts are twenty tasks, each billed at the single-image end of the scale. That loop costs around 80,000 credits, not the 35,000 that five batches of four would have suggested. If your working memory of the platform says images get cheaper in bulk, the correction is that they get cheaper within a prompt.
Put the two purchases next to each other for a real job. Twenty finished images from twenty prompts: about 80,000 credits. Four shots you have already designed, four versions each: 28,000. The second is cheaper, and for most work it is also the more useful thing to own, because what you actually needed was a choice rather than a catalogue.
What the agent is for, and where it stops paying
If variations are cheap, then the batch slider is the right tool whenever a decision is open, and an agent is the right thing to drive it. You do not want to sit there clicking generate four times and then sort files by hand. You write one prompt, ask for four, and pick.
The trap is a subtler version of the same arithmetic. Cheap variations make generating feel cheaper than deciding. A CSV of twenty prompts looks like the same gesture as a batch of four, and it is not: you have paid the full per-image price twenty times, and if you wrote those twenty lines before the look was settled, you paid it twenty times for a decision you had not made. The batch is cheap because you already know what you want. The loop is expensive because you don't.
Ai-Hakase's own closing advice is the counterweight, and it is the most practically useful line in the post: settle the model and the prompt combination in the normal web interface first, generate a handful of images by hand, and only then freeze those settings and hand them to the agent. You spend the exploratory phase somewhere you can see each result, and the agent phase on a decision you have already made.
Then, inside the run, test one row before the rest. Ai-Hakase does that: generate the first row of the CSV, look at the character, the framing, the size, check the credit burn, and only then run the remaining prompts. The loop inherits whatever the first image got wrong, so the first one is the only one worth inspecting closely.
A key is a path, not a message
The second thing worth stealing is how the credential is handled, because the setup deliberately routes around the obvious approach.
The agent is told to create a folder and a file called Sample.env.txt containing the single line PIXAI_API_KEY=. The user opens that file and pastes the key after the equals sign. Then the agent is asked to rename the file to .env and not to display its contents.
The reason given is the one that matters: anything pasted into a chat window stays in the conversation history. A file is not a transcript, and asking the agent to read it without printing it keeps the key out of the visible back-and-forth.
That is a genuine improvement on pasting a key into a message. It is not a guarantee, and the difference matters. A request not to display something is a request, not an enforced boundary. The generated script can print the environment variable into a tool result, the agent can paste it into a file it writes, and anything the script does with the value is invisible until you read the script. Treat the file as reducing exposure rather than removing it. Scope the key to the minimum it needs, rotate it after an unfamiliar run, and if the agent offers native credential storage, prefer that over a text file in a workspace the agent can read and write.
There is also a side effect the author flags honestly: once renamed, .env may vanish from the agent's file browser. It has not been deleted, it has become hidden, which is what dotfiles do. Do not ask the agent to "restore" it.
This generalises past PixAI. Any agent that needs a credential should be given a path to it, not the value. The moment a key is typed into a message, it exists in at least three places: the transcript, whatever memory the agent keeps, and the provider's logs.
The model is a version ID, not a name
The API does not take a model name. It takes a numeric modelVersionId. Ai-Hakase's method is to open the model page, pick the version, and read the trailing number off the browser URL.
That method is worth flagging, because the URL number is not always the ID the API wants. PixAI's own model reference lists them separately, and Haruka v2 shows the gap plainly: the model page sits at ...484240 while the version ID the docs tell you to send is ...989558. The reference table is the authority. A number copied out of a browser address bar is a guess that happens to be right sometimes.
I could not find a published API version ID for Tsubaki.3 in the reference table at the time of writing, so I am not quoting one. If you see a specific number anywhere, including in the walkthrough this piece is based on, check it against the table before you send it.
Which is the real point: pin the model, because "Tsubaki.3" is a label that can be repointed and the version ID is a snapshot. Record it next to the prompt, or you cannot tell six weeks later whether a change in output came from your edit or from the platform updating the model underneath you.
The same logic applies to every setting the agent freezes, plus one that people forget: the seed. The request accepts one, the response returns the resolved value back in outputs, and if you omit it the platform picks one for you. A run with no recorded seed cannot be reproduced, and you will spend an afternoon testing prompt variations that are actually testing randomness. Ai-Hakase's instruction to hold the configuration fixed across a batch is the same instinct from a different angle: a set of images is only comparable if they differ in one thing, and the prompt is the thing you meant to vary.
Where the output actually lives
One gotcha is worth knowing before you plan a library around it. Images generated through the API do not appear in PixAI's normal generation screen. The prompt history and the credit spend show up there, but the images themselves land in the agent's workspace, inside the folder the agent created. If you generate two hundred images and only then look for them in the platform's gallery, you will find nothing and assume the run failed.
The author also notes the queue, and the docs put a number on it: up to ten tasks in waiting status per account, with running tasks not counted against that cap, and the cap raisable per account on request. There is no enforced request rate limit, but the same page asks for exponential backoff on retries and polling no more often than every 1.5 seconds.
Concretely, that means the CSV loop is a queue-design problem before it is a budget problem. Thirty distinct prompts will not fire at once; they go in waves, and the script needs to know that instead of discovering it through a wall of 429s.
How to read this post
A few caveats, because the piece is not neutral. It is a PixAI-affiliated walkthrough with referral links, published days after the integration shipped, and its claims are single-source and dated to mid-November 2026. The two credit totals are one model at one setting in fast mode, and they do not establish a price list. The author's monthly credit figures come from her own login habits and will not transfer.
What survives the caveats is the shape of the thing: a chat agent plus a paid image API plus a CSV is now a working production loop for someone who cannot code, and the pricing inside that loop does not behave the way a per-image mental model predicts.
If you build one, do four things. Write down the version ID and the seed next to the prompts. Keep the key in a file rather than a message, and rotate it anyway. Use the batch slider for variations of a shot you have already decided on. And spend the care on image one, because the rest inherit its mistakes.
For the neighbouring problems, letting an LLM write the prompts inside the pipeline covers what happens when the model is in the loop for text as well, and the Tsubaki.3 panel test is the same author working through what a single image actually has to contain. If you are deciding whether the agent should hold the goal at all, the gym-booking incident is the cautionary version, and one API key across TTS, image, and transcription is the platform-side view of the same budget problem.