Hi, I'm Aimantas. I build small web products and the videos that go with them. On 5 November I bought Higgsfield's Plus plan, 1,200 credits for €59, and 56 minutes later 96.6% of them were gone.
Nobody hacked anything. Four of my working sessions, spread across three projects, reached into the same balance at the same moment, and none of them had a limit. By 20:51 I had 41 credits left and a fairly clear lesson about how AI video gets expensive.
The first step down was 512 credits in one second
The chart is the whole story in one line. At 20:09 four Seedance 2.5 clips were submitted together, each seven seconds at 1080p, and the balance dropped by 512 in the same second. Those were concept loops for a website hero: moving backgrounds nobody had picked yet.
Thirteen minutes later two lip-sync clips went through for a re-voiced YouTube video, 217 credits for 31 seconds of a talking presenter. At 20:35 another session fired off five Kling clips and one more Seedance clip in a single batch for a second website. At 20:46 someone wanted one more 1080p loop, and that was the plan done.
Each of those decisions looked reasonable inside its own session. Nobody was watching all four at once, and the balance is the only place they meet.
What €59 actually bought
Here's the honest list. The website hero concepts took 689.5 credits, about €34: fourteen still images and five seven-second video loops. The second website's hero video took 183.5, about €9. The re-voiced Transport Fever video took 269, about €13, and eight voice takes for my "Guess the flag" Shorts took 13.8, which is less than a euro.
The cheap things were the ones I actually shipped. A whole 7:34 narration in ElevenLabs, 53 lines, cost about 37 credits. That's under €2 for the entire voice track of a seven-and-a-half-minute video. Draft images were a quarter of a credit each. I'd happily pay that all day.
The expensive things were mostly questions. A concept loop for a hero section is a way of asking "does this direction feel right?" I paid finished-asset prices for answers I hadn't needed yet.
The gap between those bars is where the money went. On the same day a seven-second Seedance clip at 1080p cost me 128 credits, about €6.30, a five-second Grok Lite clip at 720p was quoted at 5. A 480p Seedance draft was 15. For a background loop that hadn't been approved, the 5-credit clip would have answered exactly the same question.
The presenter that didn't pay for itself
The one result I can measure is the Transport Fever video. I'd already published part 2 of the series with plain voice-over gameplay. For the re-voice I added a lip-synced presenter who opens and closes the video, then shrinks into a corner window while the game plays.
That corner window is 217 credits. It looked natural, honestly better than I expected. It just didn't move anything. The re-voiced upload had 41 views after three days, and it's no longer on YouTube. The original upload of the same story, plain voice over gameplay, had about 2,400 views after its first three days and 2,759 when I checked on 13 November.
That isn't a perfectly fair fight: the re-voice was a second upload of a story the channel had already told. But it's enough to say the presenter didn't earn its cost, and that a €2 narration over real footage was already the right product.
Why it happened, in three plain reasons
One balance, no budgets. Higgsfield has a single credit pool. There's no per-project limit and no spend cap on the connector, so a session working on a website hero and a session working on a YouTube video are spending from the same wallet without knowing about each other.
Defaults decide the bill. The generation skill bundled with the connector tells agents to prefer the quality default and skip cost estimates unless asked. So every busy session reached for Seedance at 1080p. Nobody chose to spend €6.30 per clip. The default did.
Batches before samples. Four Seedance clips went in one go, and five Kling clips in another. If the first clip had been reviewed before the rest were sent, most of that spend would have stopped at one.
Credits make this easy to miss. "128" feels like a small number in a tool output. "€6.30 for seven seconds of background" reads very differently, and I'd have asked whether I really needed it.
What I liked about the tool
To be fair to Higgsfield, the tool did exactly what it was asked. The connector works inside Claude Code with about 150 tools, and every cost check I ran matched the charge to the credit. ElevenLabs voice through it was quick and good: lines cost between 0.46 and 1.15 credits. When a batch hit rate limits, 8 of my 53 lines failed, and those failures weren't charged. The lip-sync quality was good enough that I shipped it.
The problem wasn't the product. It was how I let it be used.
What I changed
Every video job now starts with a price check. Then one cheap sample. Only after someone has looked at the sample does a batch run, and 1080p needs an explicit ask.
The ladder starts at Grok Lite (5 credits for five seconds) and a 480p Seedance draft (15). Lip-sync starts at 480p too: the same 12 seconds of voice quoted 39 credits at 480p against 91 at 720p. The presenter spent most of his screen time in a window about 480 pixels wide anyway.
Since then my "Guess the flag" Shorts have moved off paid generation entirely: local voice, local Blender renders, no credits at all. That's what happens when you price a thing honestly. Some of it turns out not to need the expensive tool in the first place.
Would I buy it again?
For voice and draft images, yes, without thinking. They were cheap, fast and good enough to ship.
For video, only with a limit set per job before anything runs, and only one session at a time on the balance. The lesson I'm keeping is simpler than the chart: decide what a question is worth before you ask it. A concept loop is worth 5 credits. A finished hero someone has approved might be worth 128. Paying the second price for the first job is how €59 disappears in under an hour.
The full version with the itemised table and sources is on aimantas.com. Numbers come from the account's transaction list (152 rows, read 7 November 2026) and cost quotes taken on 5 November. Project attribution comes from file times and session records, since Higgsfield doesn't tag jobs by project. Video views were read from public YouTube pages.
