HeyGen Creator Plan Credit Audit: What 439 Credits Revealed

Last updated: August 4, 2026

HeyGen credits are the unit the Creator plan meters. The HeyGen Creator plan costs $29 a month and issues 600 of them, deducted when you generate video, translate it, or create assets. We bought the plan, spent 439 of those credits across four generation modes, and examined every billing and account-record surface we located and could access during the test window. This briefing answers one question: can 600 credits be forecast before you spend them, and attributed afterward?

The pool reconciled to the credit. The per-job planning and attribution layer did not.

What 439 credits proves, and what it does not

The account’s Usage and History panel reported 439 used and 161 remaining against a 600 credit allocation. Read on 2 August 2026 at 08:17 Japan time, the history listed fourteen rows.

Thirteen of those rows carried a numeric credit value. Twelve were dated inside the current billing cycle, and those twelve sum to exactly 439. Adding the remaining balance closes the plan: 439 plus 161 is 600.

Those twelve rows correspond to ten completed outputs. The difference is the video podcast format, which is associated with two rows for each finished podcast rather than one.

Layer	Count
History rows visible	14
Rows carrying a numeric credit value	13
Rows dated inside the current cycle	12
Completed outputs those rows correspond to	10

That arithmetic is the strongest result in this audit, and it is worth stating plainly before anything else. The pool balanced. Nothing we found suggests a missing or duplicated current-cycle charge.

It also does not establish that any individual charge was correct, that a buyer could have predicted it, or that the same charge would appear on another account.

The two rows outside the twelve are worth a sentence because of where they sit. One is dated 30 July, titled after a video subject rather than a feature, and shows a value of 1 credit. The other is dated 28 July and shows no credit value at all. The subscription was purchased on 30 July at 12:38 Japan time, so both rows sit at or before the boundary between the free period and the paid cycle. The evidence is consistent with the one-credit row belonging to a pre-subscription job, and the history exposes no shared transaction identifier that would settle it either way.

The point for a buyer is narrower than a discrepancy. The visible history mixes rows from different account periods in one list, one column, and one visual treatment, with nothing on screen marking which billing cycle or credit regime applies to a given row.

Source: Future Stack Reviews hands-on session, one HeyGen Creator account accessed from Japan. Usage and History read 2 August 2026, 08:17 JST. Checkout captured 30 July 2026, 12:38 JST. Captures held by FSR and not published because they carry account identifiers. The full fourteen-row transcription is in the appendix.

HeyGen Video Agent credits: the current rate conflict

Video Agent is the prompt-driven mode. Before quoting what our jobs cost, here is what HeyGen’s own current surfaces say a minute of it costs. We opened each page and recorded the figure and the date.

Surface	Figure	Access date	Status
Japan pricing page, credit FAQ	20 credits per minute	2 Aug 2026	Official claim
Video Agent quality and unlimited mode guide	1 minute equals 20 credits, with worked examples	2 Aug 2026	Official claim
Credit-based pricing plans explained	Video Agent standard mode, 30 credits per minute	1 Aug 2026	Official claim
How to use credits on HeyGen	Video Agent standard mode, 30 credits per minute	2 Aug 2026	Official claim
How to get started with Video Agent	No credit figure stated; links to the guide above for limits	2 Aug 2026	Official claim
Template hover inside the product	About 40 per minute, shown as an approximation	1 Aug 2026	Observed by FSR

Our three paid Video Agent jobs were consistent with 40 credits per minute after whole-credit rounding. That figure appears in the table above only once, in a hover label inside the product, and none of the five official pages we opened stated it. We did not run an exhaustive search of HeyGen’s documentation, so this is an absence across the pages we checked rather than an absence across HeyGen.

One scope note belongs with those three jobs before anyone compares them to the table. All three were submitted through a connected agent rather than from the Video Agent surface inside the product, which did not accept a submission during our testing. Whether the two routes meter identically is not something this audit establishes. What it does establish is that the debits landed in the same account history, inside the same 439, on the same monthly allocation.

We are not resolving the rate question. We are not averaging the figures, and we are not nominating one of them as the correct one. The current official pages alone span 20 to 30 credits per minute, so the higher published figure is 50 percent above the lower one. The in-product approximation widens that span to about 40 credits per minute, twice the 20-credit figure. A buyer planning a month of prompt-driven video may therefore reach materially different forecasts depending on which current surface they read.

One further reading belongs here with its limits attached. FSR’s free-plan briefing recorded a signed-in tooltip stating that 600 credits covered up to 20 minutes of Video Agent. Dividing gives 30 credits per minute. That is arithmetic on a historical tooltip read on 28 July from a free account, not a displayed rate and not a paid debit.

Sources: HeyGen, Japan pricing page, read 2 August 2026 · HeyGen Help Center, 23 June 2026, read 1 August 2026 · HeyGen Help Center, 16 June 2026, read 2 August 2026 · HeyGen Help Center, 24 September 2025, read 2 August 2026 · Future Stack Reviews paid-cycle session, 1 August 2026.

The observed debit matrix

Ten completed outputs, twelve debit rows, four modes.

Mode	Outputs	Debit rows	Observed debit	What it supports	What remains unresolved
Avatar presenter	4	4	12, 13, 12, 13	Four short renders in a narrow duration band produced a consistent effective rate near 20 credits per minute	Whether the same pattern holds outside 35 to 39 seconds, and which engine executed
Video Agent	3	3	25, 29, 25	Three jobs consistent with 40 credits per minute after rounding	How that relates to the 20 and 30 figures on current official pages, and whether the in-product route meters the same way as the connected agent we used
Video Podcast	2	4	50 and 30, twice	Each completed workflow is associated with 80 credits in total	Which 50 row pairs with which 30 row, and what the 30 is calculated from
Seedance 2 video asset	1	1	150	One 1080p job matched the 150 credit figure listed on current official pages	Whether the charge varies with length, which one job cannot show

Effective rates, for the three Video Agent jobs only, using the retrieved delivered durations:

Delivered duration	Debit	Credits per delivered minute
36.2057 s	25	about 41.4
37.0939 s	25	about 40.4
43.4114 s	29	about 40.1

Two of those three jobs were submitted with byte-for-byte identical prompts of 2,115 characters. They delivered 37.0939 and 43.4114 seconds and carried debit rows of 25 and 29. Prompt text alone did not determine the runtime or the debit in those two runs. We are not offering a cause, and two runs are not a product characteristic.

We have deliberately not converted the Seedance job into a per-minute figure. Current official pages list it as a fixed charge for any length, so normalizing a four second delivery into a rate would describe our clip rather than HeyGen’s billing.

Sources: Future Stack Reviews paid-cycle session, 1 August 2026. Debits are values read from the account’s billing history. Durations are values returned by the account’s own read-only records. Seedance charge listed on HeyGen Help Center, 16 June 2026, read 2 August 2026.

What the credits delivered

Ten completed outputs across four modes. Nine delivered media files were retained and re-measured; the file-level results below therefore cover nine files, not all ten outputs. What follows is specification, not a quality score. Nothing here ranks one mode against another or against any other product.

	Avatar presenter	Video Agent, captioned download	Video Podcast	Seedance 2
Files measured	3	3	2	1
Resolution	1280×720 on two, 1920×1080 on one	1920×1080	1280×720	1920×1080
Frame rate	25	25	25	24
Video codec	h264 Main	h264 Main	h264 High	h264 High
Audio	48 kHz stereo	48 kHz stereo	44.1 kHz mono	44.1 kHz stereo
Video bitrate	1.58 to 3.85 Mbps	5.04 to 5.74 Mbps	0.85 to 0.92 Mbps	2.58 Mbps
Encoder tag	Lavc59.37.100 h264_nvenc	same	Lavc62.11.100 libx264	none present
Muxer tag	Lavf59.27.100	same	Lavf62.3.100	Lavf62.3.100

Three observations a buyer can use.

The podcast is the only mode that delivered mono audio, and the only one whose files carry a software encoder tag where the presenter and agent files carry a hardware one. Among the duration-based workflows normalized in this briefing, it carried the highest observed cost per delivered minute. Seedance is excluded from that comparison because current official pages describe it as a fixed per-job charge and this briefing does not normalize it to a per-minute rate.

The Seedance file carries no encoder tag at all, only a muxer tag, and it is the only retained file running at 24 frames per second. Those file-level differences are consistent with a different export or encoding path, but the metadata alone does not establish which generation pipeline produced it.

Both podcast files end in digital silence, 1.37 seconds on the shorter one and 1.24 on the longer, measured at a threshold of 25 dB below full scale.

One caveat belongs on the agent column. Those are the caption-burned downloads, which involve a re-encode, so their encoder tags may describe the caption pass rather than generation. We did not retrieve the versions without burned-in captions.

Source: Future Stack Reviews. All nine retained delivered media files were re-measured in a controlled environment on 2 August 2026 using ffprobe and ffmpeg. Their SHA-256 digests matched the values recorded when they were downloaded.

Presenter: selector, label, and debit did not align

The presenter composer offers an engine selector, and HeyGen publishes very different rates for the options in it. Its credit table lists Avatar III at 3 credits per minute and Avatar IV or V at 20. On a 600 credit plan that difference is the largest single lever a buyer has.

We submitted four presenter jobs. One selected Avatar III. Three selected Avatar IV. All four delivered projects displayed the label Avatar III. All four carried debits of 12 or 13 credits, an effective range of roughly 19.9 to 20.9 credits per delivered minute.

The selected engine, the delivered label, and the observed debit did not align across those four jobs.

The delivered files do not separate them either. The two 720p renders, one submitted with Avatar III selected and one with Avatar IV, are close to indistinguishable at the file level: same resolution, same frame rate, same codec profile and level, same audio layout, and the same encoder and muxer tags, with video bitrates within three percent of each other. The third file we retained differs in resolution and profile level and in nothing else we measured.

We are not claiming which engine executed. A project label is not proof of an execution path, and the account exposes no field that would settle it. We are not claiming that any charge was incorrect, that the selector failed, or that a defect exists. What we can say is that a buyer choosing Avatar III on this surface, on this account, in this window, did not see a charge in the region the published Avatar III rate implies, and had no surface available to find out why.

Sources: HeyGen Help Center, 23 June 2026, read 1 August 2026 · Future Stack Reviews paid-cycle session, four presenter renders, 1 August 2026. Composer state, delivered project label, and billing row captured for each.

Video Podcast: two stages, no join key

Two completed podcasts produced four billing rows. Two rows carry the generic feature name and a value of 50. Two rows carry the title the product generated for the finished video and a value of 30.

Before submitting, the composer displayed an estimate reading 50 credits now and 80 total, and described the charge as a fixed base at preview plus a per-minute amount at final creation. The four rows and the two completed outputs are consistent with 80 credits for each completed workflow.

The history itself exposes no shared transaction identifier linking a particular 50 row to a particular 30 row. Pairing them is an inference from ordering and value, not a record the product provides. For a buyer trying to answer what one podcast cost, that is the whole problem in one screen.

The two outputs delivered 12.520 and 45.080 seconds. Both are associated with 80 credits.

Neither of the two Help Center rate tables we checked disclosed the 50 credit preview-stage amount. Both list the video podcast at 3 credits per minute. Those two figures describe different things, an all-in workflow total against a per-minute table entry, and we are not turning the gap into a multiplier.

Sources: HeyGen Help Center, 23 June 2026, read 1 August 2026 · HeyGen Help Center, 16 June 2026, read 2 August 2026 · Future Stack Reviews paid-cycle session, two podcast workflows, 1 August 2026. In-product estimate text, all four billing rows, and both delivered files captured.

What appears before you generate

HeyGen’s credit table carries a qualifier beneath it: “When available, HeyGen will show the credit cost in-product before you generate.” We recorded what “when available” resolved to on the surfaces we used.

Surface	Estimate shown before submit	Relationship to the debit
Video podcast composer	Yes, with a stated two-part model and a settings-linked figure	Consistent with the rows that followed
Video Agent template hover	Yes, an approximate per-minute figure	Consistent with the three observed jobs
Seedance tooltip	Yes, a static 60 that did not respond to the resolution control	The 1080p job carried 150
Presenter composer	No, across four separate conditions we checked	Not applicable
Video Agent free prompt entry	No	Not applicable
Clip generation	No, and the submit control remained active	Not applicable
In-product assistant, asked directly	No, across four engine and duration questions	Not applicable

Three of the six tested generation surfaces disclosed a pre-submit estimate, and two of those estimates were consistent with the debit that followed. Across this six-surface sample, we did not identify a reliable relationship between estimate availability and feature age or mode.

There is a seventh surface, and it is the one a confused buyer reaches for first. We asked the in-product assistant what a minute costs on each avatar engine and it declined every time, redirecting to Plan and Billing. In the same session it described the Creator plan as having an unlimited monthly video allowance and, two answers later, reported 600 credits ready for the cycle. It also stated that Creator credits do not roll over. The pricing page, both credit tables, and the account’s own cancellation dialog all state the opposite.

Sources: HeyGen Help Center, 23 June 2026, note beneath the feature credit table, read 1 August 2026 · Future Stack Reviews session, six composer and tooltip surfaces, 30 July to 1 August 2026.

Three record surfaces, no common job-cost key

After a charge lands, a Creator subscriber has three places to look. We read all three.

	Usage and History	CSV export	Read-only account records
Credit value	On 13 of 14 rows	On 4 of the 7 rows in the export	No per-job debit field in the records we retrieved
Duration	On 8 of 14 rows	On 6 of 7 rows	Present, to four decimal places
Time of day	Not shown, date only	Not shown, date only	Present, to the second
Which feature was billed	Mixed: feature names and generated video titles	Same mixture	Job type present, cost absent

Among the three record surfaces retrieved in this audit, we found no per-job identifier that appeared on more than one surface. That is the operative finding for this account and test window. Each surface held part of what a reconciliation needed, and we found no field that joined all three.

Two further observations sit under that.

The project column mixes feature names with generated video titles, so a row reading like a video headline cannot be traced back to a feature from the billing screen alone.

And the export’s duration column does not consistently describe the delivered file. One export row reads 37.4 seconds, while the corresponding retained-media duration does not produce 37.4 under conventional rounding. Two other rows read 35.2 against a file measuring 35.202 seconds, and 4.0 against a file measuring 4.06 seconds. We did not identify a single rounding rule that reconciled every export value to the delivered media. The export values were, however, consistent with the shorter durations returned by the account’s read-only records. The evidence therefore supports a narrower conclusion: the export and the downloaded media describe different duration measures. A buyer reconciling the export against delivered files should not assume that the two columns are directly interchangeable.

The export also carries a member column populated with an account identifier. It is the only one of the three surfaces that does.

Source: Future Stack Reviews session, one Creator account. Billing history read 2 August 2026, CSV exported 1 August 2026 at 14:25 and 14:40 JST, read-only account records retrieved 1 August 2026. Delivered media measured locally. The export predates the evening jobs, so rows created afterward are absent from it for that reason.

Concurrency contaminated one attribution

Because none of the records retrieved in this audit exposed a per-job cost field, balance change was the only automated attribution method available to our connected workflow in this test.

In one observation window that method returned a wrong number. A balance change of 38 credits spanned a 25 credit Video Agent debit and a separate 13 credit browser debit submitted in the same period. Balance-delta attribution conflated the two, and nothing in the sequence signaled that it had.

This is one observation, not a rule. It was enough for us to reject balance subtraction as our primary attribution method when jobs could overlap on the tested account. It does not establish how every account, integration, or automation behaves.

Source: Future Stack Reviews paid-cycle session, 1 August 2026. Both debits later appeared as separate rows in the account billing history, which is how the conflation was identified.

What the Japanese checkout displayed

HeyGen’s Japanese-language pricing page quotes every plan in US dollars and shows no yen figure. The checkout screen does, and it was the only surface in this test that did.

Line	Displayed
Product	HeyGen Creator Platform, billed every one month
Currency control	Two buttons, JPY and USD. JPY was selected in our capture
Disclosed conversion	1 USD = 170.0183 JPY, with a note that the billed amount moves with the exchange rate
Subtotal	¥4,931($29.00)
Tax	¥0
Total due today	¥4,931($29.00)

The arithmetic checks: 29 multiplied by 170.0183 is 4,930.53, which rounds to the displayed ¥4,931.

The checkout let us choose between JPY and USD. In the capture used for this briefing, JPY was selected and the checkout displayed ¥4,931 alongside a $29.00 reference amount. We completed the purchase in yen. The subscription management portal afterward listed the plan in US dollars while noting that the charge was made in JPY.

What we have not done is open a card statement. So the exact settled amount, whether the issuer applied any handling of its own, the merchant of record, and the final tax treatment are recorded as displayed rather than as confirmed. The yen figure also appears in only one place going forward: the plan and billing screen inside the product shows no amount at all, and the management portal returns to dollars.

For sizing individual jobs, the honest denominator is the allocation rather than a currency conversion. The Seedance render consumed 150 of 600 credits, a quarter of the month in one submission. Each completed podcast consumed 80, or about 13 percent. The whole test consumed 439, about 73 percent.

Sources: HeyGen, Japan pricing page, read 2 August 2026 · Future Stack Reviews checkout capture, 30 July 2026, 12:38 JST, one account accessed from Japan.

Which workflows this evidence supports

The findings above map onto procurement decisions unevenly. Some workflows are fine on this evidence. Others are not supported by it, which is different from being shown to fail.

Workflow	Decision	Reason	Evidence boundary
Solo operator, one mode, manually watched	Suitable with manual reconciliation	The pool reconciled exactly, and the history carries per-row values	One account, one cycle
Monthly forecasting across several modes	High friction	Current official pages disagree on at least one mode’s rate, and two of four modes we were billed for carry no rate on the pre-purchase pricing page	Four modes tested
Two people generating at once	Attribution risk observed	Balance subtraction conflated two overlapping debits in one window	One observation
Agency or client chargeback	Not supported by the tested records	No per-job identifier appears on more than one record surface	Three surfaces checked
Automation needing a preflight quote	Not supported by this audit	No retrieved record exposed a per-job cost field	Records we retrieved
Scoring output quality	Outside scope	Files were measured, but no rubric was set and nothing was compared against another product	Delivered specifications only

The upgrade and billing boundary

Two boundaries matter for a buyer sizing this plan.

The first is what happens when 600 runs out. HeyGen states that Creator does not support one-time credit purchases and that the route to more credits is an upgrade to Pro, which its pricing page lists from $49 a month. Running dry mid-cycle is a plan change rather than a top-up, so the practical advice is to end the month above zero rather than at it. This is a current official claim and worth rechecking, since plan mechanics have changed more than once.

The second is which balance an integration spends. The three Video Agent jobs in this audit were submitted through a connected agent rather than the browser, and their debits appear in the same account history and inside the same 439. On this account, automation and hand work drew on one pool. HeyGen’s API product is separate from that: it is a standalone pay-as-you-go purchase billed in US dollars rather than a subscription, and we did not test it. FSR’s pricing briefing covers the documented split between the two.

The same billing screen offers a control to pause the subscription rather than cancel it. It is not a cancellation, and it appears in none of the pages we opened for this briefing.

The cancellation dialog is the more informative of the two. It states the downgrade date, that payments and add-on renewals stop automatically, and that the remaining plan stays active to the end of the cycle, and it restates the rollover rule that the in-product assistant had contradicted a day earlier.

Sources: HeyGen Help Center, 16 June 2026, read 2 August 2026 · HeyGen Help Center, API pricing, read 2 August 2026 · HeyGen, Japan pricing page, read 2 August 2026 · Future Stack Reviews paid-cycle session, 1 August 2026.

Limitations

One account, one region, one cycle. Everything here comes from a single Creator subscription purchased in Japan on 30 July 2026 and observed through 2 August. No second account, region, or billing cycle was tested, and any observation could be specific to this one.

No overbilling is established. Every job completed. The current-cycle total reconciled exactly. Several charges matched an official or in-product figure. The differences we report may be specific to this account, this rollout, this mode, or this surface.

Delivered files were measured, but no quality verdict is offered. We downloaded and inspected the delivered media and recorded resolution, duration, frame rate, codec, and audio channel layout for each retained file, along with on-screen text and shot counts on specific renders. What this briefing does not contain is a scored assessment of avatar realism, voice, or lip sync, and no comparison against HeyGen’s other engines or any competitor. Those require a rubric this audit did not set.

Delivered resolution varied inside the test window and we did not resolve why. Three presenter renders submitted in the morning delivered at 1280 by 720 after a 1080p selection. One submitted the same evening delivered at 1920 by 1080. At least three explanations survive, including a vendor-side change, a resolution control at the download step, and a difference in how the file was retrieved. We did not separate them, so this is recorded as an open item rather than a finding.

Untested on this account. 1080p video podcast pricing, the 720p Seedance charge, digital twin training, and any output longer than 46 seconds.

Cancelled but not yet observed. We cancelled the subscription on 2 August. The plan does not downgrade until 30 August, so what happens to the 161 remaining credits, and whether the watermark returns to videos already made under the paid plan, are outside this briefing. The screen that confirmed the cancellation called 30 August a downgrade. The plan screen, six minutes later, still called the same date a renewal.

Three items remain open. Which 50 credit podcast row pairs with which 30 credit row. Whether the one-credit row dated 30 July belongs to the free period. And which of the two descriptions of 30 August the account will act on. None can be settled from what the account exposes today.

Rights are plan-scoped and we are not reading them for you. HeyGen’s terms address paid-plan output in one section and free-plan output in another, and the two are drafted differently. Anyone whose purchase depends on output rights should read the current terms and take their own advice.

Sources: HeyGen, Terms of Service, last updated 23 July 2026, read 1 August 2026 · Future Stack Reviews session, 28 July to 2 August 2026.

FAQ

Methodology and source note

This is a Tier B briefing. Future Stack Reviews purchased a HeyGen Creator subscription on 30 July 2026 at 12:38 Japan time and ran structured paid tests on 1 and 2 August 2026, from one account accessed from Japan on desktop Chrome.

Independent re-verification. Nine retained delivered media files and both CSV exports were rechecked on 2 August 2026 in a controlled environment. All eleven SHA-256 digests—nine media files and two CSV exports—matched the values recorded at download, and file sizes matched to the byte. Durations, resolutions, codecs, audio layouts, frame counts, and encoder tags reported for the retained media files come from that re-measurement rather than from a session note. One earlier observation could not be reproduced: a blink count recorded in a prior session was re-attempted with a different detection method, which measures a different thing, so it is not reported here.

What was measured. Ten completed outputs across four modes produced twelve current-cycle debit rows. Nine delivered media files were retained for file-level measurement. For each job we recorded the composer state before submission, the billing row afterward, and the delivered file where one was retrieved. File runtimes were measured locally from the downloaded media. The account history was transcribed by hand from screen captures taken on 2 August at 08:17 Japan time, across both pages of the paginated list.

Why balance subtraction was not our primary method. No record we retrieved exposed a per-job cost field, so balance change is the only automated measurement available. We used it early, found it conflated two overlapping debits in one window, and moved to reading per-row values from the billing history instead. Every credit figure in this briefing is a row value, not a subtraction.

Where linkage remains inferential. Pairing a 50 credit podcast row with a 30 credit row, and assigning the one-credit row dated 30 July to the free period, both rest on ordering, value, and date rather than on an identifier the product provides. Both are marked as unresolved in the body.

How official sources were treated. Every vendor page cited was opened directly at the URL shown, on the date shown. A vendor page supports a claim about what that page stated on that date. It does not verify how any account is billed. Where pages disagree, we record the disagreement rather than choosing between them.

What the sibling briefings contributed. FSR’s pricing briefing recorded public page states on 28 July from a signed-out session, and its figures are dated to that reading rather than treated as current. FSR’s free-plan briefing recorded free-account behavior on 28 and 30 July. Neither supports any paid-cycle observation here. No claim from either was carried forward without being independently established in this audit.

On the previous version of this page. This URL previously carried a HeyGen workflow tutorial published in March 2026. It has been replaced in full. None of its figures, rates, or plan descriptions were carried forward or used as evidence, and several were contradicted by the measurements above.

Evidence handling. Screen captures, exports, delivered media, and hashes are held by Future Stack Reviews and are not published, because they carry account identifiers. That is why this briefing carries no numbered references. Each section instead names the surface, date, and time an observation came from.

Measured but not judged. Delivered resolution, duration, frame rate, codec, and audio channel layout were recorded for every retained file. No quality score, no competitor comparison, and no assessment of support, refunds, or any legal question about the terms appears here.

Verdict

At the level of the pool, this plan behaved. Six hundred credits arrived, twelve current-cycle rows reconciled to 439, the balance closed to the credit, and every job we paid for produced a row we could point at. That is more than the audit set out expecting.

One level down, where buyers actually plan, it did not. A minute of prompt-driven video carries two different published figures depending on which HeyGen page you read, and our jobs matched neither. An engine selector that carries a large published price difference produced no difference in what we were charged. Half the surfaces we submitted from showed no cost estimate at all. And once a charge lands, the screen with the money, the export, and the machine-readable records each hold a different part of what a reconciliation needs, with no shared key between them.

Future Stack Reviews paid ¥4,931 for the account in this briefing and earns nothing from HeyGen. This page carries no affiliate link, no referral code, and no tracking parameter. If you decide to subscribe, go to HeyGen directly.

Last checked and what changed

Item	Status
Official rate pages	Read 1 and 2 August 2026
Japan pricing page	Read 2 August 2026
Account balance and history	Read 2 August 2026, 08:17 JST
Terms of Service	Read 2 August 2026. Last Updated date unchanged at 23 July 2026
Delivered files and exports	Re-measured and hash-verified 2 August 2026
What changed in this update	First publication of the paid-cycle audit at this URL, replacing the March 2026 workflow tutorial in full

Appendix: the fourteen-row history exhibit

Transcribed from the Usage and History panel of one HeyGen Creator account, read 2 August 2026 at 08:17 Japan time, across both pages of the paginated list. Rows 1 to 12 are dated 1 August and sum to 439, the figure the same panel displayed as used. Rows 13 and 14 are dated earlier and are not part of that total.

#	Project label	Date	Duration shown	Credits
1	Future Stack Reviews: The Cost of AI	1 Aug 2026	—	25
2	Avatar Video	1 Aug 2026	00:39	13
3	Future Stack: AI Audit Report	1 Aug 2026	—	29
4	Future Stack Audit: The Receipts	1 Aug 2026	—	25
5	Keeping the Receipts on AI Costs	1 Aug 2026	00:46	30
6	Podcast	1 Aug 2026	—	50
7	Test Sequence Four One Seven	1 Aug 2026	00:13	30
8	Podcast	1 Aug 2026	—	50
9	Seedance 2.0 Video	1 Aug 2026	00:04	150
10	Avatar Video	1 Aug 2026	00:37	12
11	Avatar Video	1 Aug 2026	00:38	13
12	Avatar Video	1 Aug 2026	00:36	12
13	The Origin of Twitch	30 Jul 2026	—	1
14	Avatar Video	28 Jul 2026	00:08	—

The dash character in the duration and credit columns is the product’s own display for an empty value and is transcribed as shown.

Editorial correction, 4 August 2026. We clarified that the current official Video Agent pages differ by 50 percent rather than by a factor of two; reconciled ten completed outputs with nine retained media files and two CSV exports; narrowed the per-minute, pipeline, record-surface, and pause-control language to the tested evidence; and corrected the JPY/USD checkout description. No credit debit, account-balance total, or workflow verdict changed.