274 Higgsfield Credits in One Week: Per-Clip Charges and Spend the Guard Missed
This post contains a Higgsfield referral link; we may earn a commission if you subscribe.
Corrected version. First published October 7, 2026. We re-checked every number against our credit log afterwards and changed this:
- The price table now counts charged calls (46) instead of quoted requests, and the three jobs our guard never saw are listed separately.
- We removed the sentence saying every connector goes through the same guard. It does not. See “What the guard covers”.
- We removed “charges that land after the call returns”. In this week’s log no charge landed late.
- The default GPT Image 2.5 calls were 3:2, not 9:16 as the first version implied.
- The pointer to a data file is gone. The tables below, plus the appendix with every request, are the data.
- New section with numbers collected on October 13, 2026, after the account moved to the Max plan.
We make short videos with a small team of AI agents. Most paid generations go through a guard script that writes three rows to a ledger: the price quoted before the call, the balance just before it, and the balance just after it. Between September 28 (11:12 KST) and October 4 (12:06 KST) our Higgsfield balance went from 986.46 to 712.21 credits. That is 274.25 credits in a little over six days. This post is what the ledger says about where those credits went, what each clip type cost, and which part of the drop our own guard did not see.
The nine unchanged balance observations did not prove delayed charges. They consisted of three requests that returned preset suggestions without a job, three reference uploads and three reference confirmations. In the 46 charged calls, the quote and measured spend matched.
Where the 274.25 credits went
| Day (KST) | Charged calls | Credits | Note |
|---|---|---|---|
| Sep 28–29 | 0 | 0 | Ledger started at 986.46 on Sep 28, 11:12 |
| Sep 30 | 11 | 42.25 | Kids story keyframes and shot tests |
| Oct 1 | 3 | 24 | Song video shots |
| Oct 2 | not in the ledger | 61.75 | Three jobs the guard did not see (table below) |
| Oct 3 | 25 | 66.25 | Kids story re-takes, character sheets, two 3D tests |
| Oct 4 | 7 | 80 | 60-second kids story clips and one comedy draft, until 12:06 |
| Total | 46 + 3 jobs | 274.25 | 986.46 − 212.5 − 61.75 = 712.21 |
By type: video took 205.75 credits, 3D models 36, stills 32.5. Of the 274.25, the guard measured 212.5 across 46 charged calls and missed 61.75.
Price per call, as quoted on the day
These are the credit prices quoted right before each call, taken from the provider’s cost estimate and typed into the ledger by the agent making the call. Prices can change, so check before you budget. For each of the 46 charged calls, the recorded balance decrease matched the quotation. The interval between the pre-call and post-call ledger timestamps was a median of 3 seconds and a maximum of 25 seconds; this measures record timing, not a delay after the provider returned.
| What we generated | Quote per call (credits) | Per second of video | Charged calls | Credits charged |
|---|---|---|---|---|
| Image, GPT Image 2.5, default quality, 3:2 | 0.25 | – | 15 | 3.75 |
| Image, GPT Image 2.5, high quality, 2K, 2:3 | 2.75 | – | 1 | 2.75 |
| Image, Nano Banana Pro, 2K, 9:16 | 2 | – | 13 | 26 |
| Video, Kling 2.6, 5 s, 9:16 | 5 | 1 | 1 (3 requests, 2 never charged) | 5 |
| Video, Kling 2.6, 10 s, 9:16 | 10 | 1 | 5 | 50 |
| Video, Kling 3.0 standard, 5 s, 9:16 | 7.5 | 1.5 | 3 | 22.5 |
| Video, Kling 3.0 standard, 6 s, 9:16 | 9 | 1.5 | 1 | 9 |
| Video, Kling 3.0 standard, 15 s, 9:16 | 22.5 | 1.5 | 1 | 22.5 |
| Video, Kling 3.0 pro, 5 s, 9:16 | 8.75 | 1.75 | 4 (6 requests, 2 never charged) | 35 |
| Image(s) to 3D model, Tripo H3.1 | 18 | – | 2 | 36 |
| Total | 46 | 212.5 |
Where our quote notes mention sound, it was off. Upload and confirm steps with no observed balance decrease are not counted: six of them, recorded with a nominal 0.01 estimate, moved the balance by 0.
The practical reading: one 5-second Kling 3.0 standard clip (7.5) costs the same as thirty default GPT Image 2.5 stills (0.25 each). For our kids story we draw the keyframes as stills first and paid 5 to 8.75 credits for the five-second attempts listed here. The ten-second Kling 2.6 attempts cost 10 credits each; these are historical observations for their stated durations.
Spend the guard did not see: three jobs on October 2
These three jobs are in the provider’s job records and in the balance trail, but not in the guard’s own rows. Each charged exactly its quoted price, and nothing else moved the balance between them.
| Job (Oct 2, 10:03–11:49 KST) | Model | Length | Credits | Balance |
|---|---|---|---|---|
| Paper plane, opening shot | Seedance 2.5, 720p | 5 s | 35 | 920.21 → 885.21 |
| Paper plane, pick-up shot | Kling 3.0 standard | 5 s | 7.5 | 885.21 → 877.71 |
| Paper plane, release and landing | Wan 3.0 | 11 s | 19.25 | 877.71 → 858.46 |
The guard’s balance check flagged the missing 61.75 at 09:55 on October 3, about 22 hours after the last of the three jobs, and we had it traced to these job records by 14:16. Two of the three job records name the assistant without an automatic hook as the maker; the third job record names no maker. For those 22 hours the guard’s day total was lower than what we had really spent.
What the guard covers, and what it does not
The first version of this post said every connector that can spend credits now goes through the same quote → pre → post rows. That overstated it, and the ledger shows why.
- Covered: calls made from Claude Code sessions. A hook runs before each Higgsfield call, requires a fresh quote and checks our own per-job, daily and monthly caps, then reads the balance again afterwards.
- Not covered automatically: the assistant without a hook, claude.ai chats, and anything done on Higgsfield’s own website. That assistant is told to run the guard by hand and record the quote before each generation. In this week 7 of the 46 charged calls (80 of 212.5 credits) were logged that way.
- Unseen spend still counts: when the guard finds an unexplained drop at the next balance check it books it as that day’s usage, so the daily cap counts the balance drop and not only the calls we saw. It is counted after the fact, though.
It happened again after this week. Between October 4 (12:06) and October 8 (02:09) five unattributed balance observations are grouped into the two windows below:
| Found (KST) | Unattributed drop | What the ledger shows |
|---|---|---|
| Oct 7, 21:35 | 30 | No rows at all for the interval since Oct 4, 12:06. The only trace is one 20-credit quote for a 5-second Seedance character test scene at 20:03 that day, with no pre or post row. The ledger cannot itemise the 30. |
| Oct 8, 02:00–02:09 | 8.24 unattributed, within an 8.25 balance decrease | After accounting for an earlier recorded 0.01 charge, 8.24 remained unattributed. Four later requests had quotation rows without pre-call or post-call balance records: three images quoted at 0.25 and one five-second Kling 3.0 Turbo clip quoted at 7.5. Those quotations alone do not identify which requests caused the balance decrease. |
That is 38.24 credits in total. From October 8, 02:09 through the last ledger row we checked (October 13, 22:01 KST) the ledger shows no further unexplained drop. The balance increase observed after the owner switched plans on October 13 is described separately below; its grant or refund composition was not itemised.
Calls that created nothing
Four generation requests in the week were quoted but never charged: 27.5 credits of quotes that did not become spend. All four were video.
- Three times a video request came back with only a suggestion to use a preset and no job (September 30 at 16:08, October 3 at 14:04 and October 3 at 18:03). The fix is to send the same request again with the preset declined. On September 30 and on the evening of October 3 that repeat ran 22 and 16 seconds later.
- The repeat of the October 3 14:04 request, sent at 14:05, was blocked inside our own agent session and never reached Higgsfield. That row has a pre entry but no post entry. The shot finally ran at 16:05, after the owner told us directly to go ahead.
If you script generations, check that you actually got a job ID before you count a clip as started. Our post rows did not record job IDs this week; the first rows that do are from October 8.
Rows where the balance did not move right after the call
The first version said that in 9 of 55 calls the charge showed up later. The ledger does not support that. Nine of the 55 post rows showed no balance change: six were upload and confirm steps with an unchanged recorded balance, and three were the preset-only requests above. The records did not establish a later charge for any of those nine observations. Every balance in the week is explained by the 46 charged calls plus the three October 2 jobs. The guard still flags a no-change call as pending and re-checks it at the next balance reading, in case a charge does arrive late, but the reviewed window did not establish such a charge.
What we regenerated, and why
Our quote notes flag three paid re-dos after a reviewer caught a visible defect: a towel that changed colour between frames (8.75 credits), fake Korean lettering on a shop sign and poster in a song-video shot (7.5), and a random defect in a still (2). That is 18.25 credits within the 274.25-credit window, and 3 of the 46 charged calls.
The same 5-second kids-story shot was tried three times in about 18 minutes on September 30 (8.75 each, 26.25 credits) and failed review every time: the first had the towel change colour and characters leave the frame, and the second and the redesigned third each had a character leave the frame. We stopped re-rolling it and rebuilt the keyframes; the next attempt, on October 3, started from new keyframes.
Limits of this test
- One account, one week, our own prompts. Your mix of images and video will give a different total.
- Prices are the quotations recorded for this account on those dates. The owner record identifies Plus as the previous plan before the October 13 switch to Max; it does not establish a general price comparison across plans. We did not score visual quality across models here.
- The 274.25-credit total includes pilots we later dropped, so it is not “cost per published video”.
- Quotes are typed in by the agent making the call. They matched the charge on all 46 charged calls, but the guard cannot verify a quote before the call. If a charge ever exceeds its quote, the guard locks further generation until the owner clears it.
- What the three preset-only responses said is in our session notes, not in the ledger rows themselves.
After the plan change: Max, records through October 13, 2026 (22:01 KST)
This section is a separate dataset, a different plan and a different model, so do not add its totals to the week above.
- Plan and balance: on October 13 the account owner moved the account from Plus to Max. Higgsfield’s balance endpoint reported plan type “max” and 1,800 credits at 09:51 (KST). The last balance we had read before the switch, at 08:53, was 379.84.
- Our ledger recorded an estimated explanation: at 10:49 the guard logged the jump of 1,420.16 credits with its default note for any balance increase, “refund or grant (estimated)”. The increase was observed after the owner switched to Max. The balance responses establish the reported plan and balance, but do not itemise how much came from a grant, refund or other adjustment. The guard does not reduce recorded usage because of an increase, and as of this writing it still writes that same default note.
- Same quote, same charge: a 10-second 720p Seedance 2.0 Mini clip was quoted 10 credits at 08:53 on October 13 before the switch and 10 credits at 10:49 after the switch. The matched before-and-after comparison is limited to the ten-second 720p quotation of 10 credits. Earlier 480p quotations were 0.5 credits per second, but the checked records contain no post-switch 480p quotation, so they do not establish a before-and-after 480p comparison.
| Request (Oct 13, KST) | What | Quote | Charged | Balance after |
|---|---|---|---|---|
| 10:49 | Seedance 2.0 Mini, 10 s, 720p, 9:16 | 10 | 10 | 1,790 |
| 11:52 | Seedance 2.0 Mini, 15 s, 720p, 9:16 | 15 | 15 | 1,775 |
| 11:55 | Seedance 2.0 Mini, 15 s, 720p, 9:16 | 15 | 15 | 1,760 |
| 22:00 | Seedance 2.0 Mini, 7 s, 720p, 9:16 | 7 | 7 | 1,753 |
| Total | 4 submitted provider requests, 4 jobs, 0 provider refusals | 47 | 47 | 1,753 at 22:01 |
Earlier on October 13, before the switch, the same 10-second request was refused with the reason code grace_period_daily_generation_limit even though the balance was far above the quote. That day is written up in Balance 379 credits, quote 10, request refused. The four submitted requests in this table were accepted, but this does not show that Max removes every limit. A separate post-switch attempt was stopped by the local guard before provider submission. These outcomes must remain distinct, and the submitted sample covers only this one model.
We do not reproduce Higgsfield’s plan prices or monthly credit amounts here, because they are the provider’s to state and they change. Check the pricing page before you decide.
Should you pay for it?
The recorded still-image calls totalled 32.5 credits within this historical window. The table describes the observed prompts and charges for this account; it is not a current price quote or a quality comparison. Check the provider’s current quotation and generation eligibility before a new request. If you choose to try the service, this is our referral link: Higgsfield.
Related tests from our own pipeline: balance 379 credits, quote 10, request refused: one day of video-credit logs, VoxCPM2 vs ElevenLabs on the same 30-second script and same footage, different file: how we checked for duplicates frame by frame.
Sources checked
- Our own guard ledger, September 28 to October 13, 2026. The guarded totals are sums of those rows, while the three October 2 jobs are reconciled from the separate job and balance records; the appendix lists 50 generation requests recorded by our guard; the three October 2 jobs outside the guard are listed separately above.
- Provider job records for the three October 2 jobs, and the owner record identifying the previous plan as Plus, and the October 13 provider balance response at 09:51 (plan type max, 1,800 credits).
- Higgsfield pricing page. We quote no prices from it.
Appendix: 50 generation requests recorded by our guard
All times are KST. “Balance” is the recorded post-call balance where a post-call reading exists; a pre-call-only refusal is marked explicitly. Six upload and confirm steps with an unchanged recorded balance are left out. This table lists 50 generation requests recorded by our guard: 46 charged calls, three preset-only responses and one locally blocked request. The three October 2 jobs outside the guard are listed separately above.
| # | Time | Request | Quote | Charged | Balance | Note |
|---|---|---|---|---|---|---|
| 1 | Sep 30, 09:02 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 984.46 | kids story |
| 2 | Sep 30, 09:04 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 982.46 | kids story |
| 3 | Sep 30, 11:20 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 980.46 | kids story |
| 4 | Sep 30, 11:41 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 978.46 | kids story |
| 5 | Sep 30, 12:27 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 976.46 | kids story |
| 6 | Sep 30, 12:29 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 974.46 | kids story |
| 7 | Sep 30, 12:44 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 972.46 | kids story. Re-do of the previous still: a second reviewer flagged a random defect. |
| 8 | Sep 30, 16:03 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 970.46 | kids story |
| 9 | Sep 30, 16:08 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 0 | 970.46 | kids story. No charge. Preset suggestion only, no job; same request sent again 22 s later with the preset declined. |
| 10 | Sep 30, 16:09 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 8.75 | 961.71 | kids story. Failed review: towel changed colour between frames and both characters left the frame. |
| 11 | Sep 30, 16:16 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 8.75 | 952.96 | kids story. Re-do after that failure. Failed review: a character left the frame. |
| 12 | Sep 30, 16:26 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 8.75 | 944.21 | kids story. Redesigned attempt at the same shot. Failed review: a character left the frame with the bundle. |
| 13 | Oct 01, 18:05 | Kling 3.0 standard, 6 s, 9:16 | 9 | 9 | 935.21 | song video |
| 14 | Oct 01, 18:06 | Kling 3.0 standard, 5 s, 9:16 | 7.5 | 7.5 | 927.71 | song video |
| 15 | Oct 01, 18:11 | Kling 3.0 standard, 5 s, 9:16 | 7.5 | 7.5 | 920.21 | song video. Re-do: reviewer flagged fake Korean lettering on a shop sign and poster; we asked for a close-up with no signs. |
| 16 | Oct 03, 10:06 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 856.46 | kids story |
| 17 | Oct 03, 14:04 | Kling 2.6, 5 s, 9:16 | 5 | 0 | 856.46 | kids story. No charge. Preset suggestion only, no job. |
| 18 | Oct 03, 14:05 | Kling 2.6, 5 s, 9:16 | 5 | not submitted; no recorded charge | not read after refusal | kids story. No charge. Blocked inside our own agent session before it reached Higgsfield; no post row was written. |
| 19 | Oct 03, 16:05 | Kling 2.6, 5 s, 9:16 | 5 | 5 | 851.46 | kids story |
| 20 | Oct 03, 16:24 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 849.46 | kids story |
| 21 | Oct 03, 16:28 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 847.46 | kids story |
| 22 | Oct 03, 17:53 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 845.46 | kids story |
| 23 | Oct 03, 17:57 | Nano Banana Pro, 2K, 9:16 | 2 | 2 | 843.46 | kids story |
| 24 | Oct 03, 18:03 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 0 | 843.46 | kids story. No charge. Preset suggestion only, no job; same request sent again 16 s later. |
| 25 | Oct 03, 18:03 | Kling 3.0 pro, 5 s, 9:16 | 8.75 | 8.75 | 834.71 | kids story |
| 26 | Oct 03, 20:01 | 3D, Tripo H3.1 | 18 | 18 | 816.71 | game character art |
| 27 | Oct 03, 21:19 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 816.46 | game character art |
| 28 | Oct 03, 21:19 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 816.21 | game character art |
| 29 | Oct 03, 21:19 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 815.96 | game character art |
| 30 | Oct 03, 21:19 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 815.71 | game character art |
| 31 | Oct 03, 21:19 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 815.46 | game character art |
| 32 | Oct 03, 21:20 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 815.21 | game character art |
| 33 | Oct 03, 21:20 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 814.96 | game character art |
| 34 | Oct 03, 21:21 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 814.71 | game character art |
| 35 | Oct 03, 21:21 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 814.46 | game character art |
| 36 | Oct 03, 21:21 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 814.21 | game character art |
| 37 | Oct 03, 21:21 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 813.96 | game character art |
| 38 | Oct 03, 21:21 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 813.71 | game character art |
| 39 | Oct 03, 21:22 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 813.46 | game character art |
| 40 | Oct 03, 21:22 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 813.21 | game character art |
| 41 | Oct 03, 21:22 | GPT Image 2.5, default quality, 3:2 | 0.25 | 0.25 | 812.96 | game character art |
| 42 | Oct 03, 21:24 | GPT Image 2.5, high quality, 2K, 2:3 | 2.75 | 2.75 | 810.21 | game character art |
| 43 | Oct 03, 21:25 | 3D, Tripo H3.1 | 18 | 18 | 792.21 | game character art |
| 44 | Oct 04, 02:11 | Kling 3.0 standard, 5 s, 9:16 | 7.5 | 7.5 | 784.71 | kids story |
| 45 | Oct 04, 02:47 | Kling 2.6, 10 s, 9:16 | 10 | 10 | 774.71 | kids story |
| 46 | Oct 04, 02:47 | Kling 2.6, 10 s, 9:16 | 10 | 10 | 764.71 | kids story |
| 47 | Oct 04, 02:47 | Kling 2.6, 10 s, 9:16 | 10 | 10 | 754.71 | kids story |
| 48 | Oct 04, 02:47 | Kling 2.6, 10 s, 9:16 | 10 | 10 | 744.71 | kids story |
| 49 | Oct 04, 02:57 | Kling 2.6, 10 s, 9:16 | 10 | 10 | 734.71 | kids story |
| 50 | Oct 04, 12:06 | Kling 3.0 standard, 15 s, 9:16 | 22.5 | 22.5 | 712.21 | comedy short draft |
Comments
Post a Comment