← Home
CAPABILITY

Published · Updated

AI Video Editing for Long Videos

Valmera edits a single uploaded video up to 14 GB or 3 hours long — whichever you hit first — in MP4, MOV, MKV or WebM. The file is indexed once into a word-level transcript with speaker labels, measured silences, shot boundaries and labeled frame tiles the agent reads directly; every request after that reads the index rather than the file. Indexing costs no credits. The export is rendered from your original upload at source quality, however long it is.

Most tools in this category cap uploads in minutes, and most do not publish the cap anywhere you can find it before you have signed up. The section below tabulates what eight of them actually state, with sources. Then the part that matters more than the ceiling: why long recordings are genuinely hard for an AI editor, what changes at the one-hour and three-hour marks, and a worked plan for a two-hour podcast.

Upload the Whole Recording

14 GB or 3 hours per file, on every plan including free. 50 credits on signup, no card.

Start free →
See pricing →

Two Ceilings: 14 GB and 3 Hours

There are two limits and both have to hold. Duration is capped at 3 hours because duration is what costs machine time — it drives how long the file takes to index and how long every render takes. Size is capped at 14 GB for a different reason: bytes cost storage and your own upload time, so that number only has to be big enough to stop being the thing that says no. Both are checked before the upload starts, not after, because refusing a file at the end of a 40-minute upload is not a kindness.

Valmera's two upload ceilings and where common recordings sitTwo scales. The duration scale runs from zero to three hours and stops at three hours; a 45-minute webinar, a 2-hour podcast and a 3-hour stream all fit, and a 4-hour stream is refused. The file size scale runs from zero to 14 gigabytes; the 2 gigabyte cap this product used to enforce sits at roughly one seventh of the current one, a 2-hour 1080p recording at 10 Mbps is about 8.4 gigabytes, one hour of 4K at 30 Mbps is about 12.6 gigabytes, and one hour of 4K60 at 50 Mbps is about 22 gigabytes and is refused.DURATION · CEILING 3 HOURSFILE SIZE · CEILING 14 GBREFUSED01h2h3h45-min webinar2-hour podcast3-hour stream4-hour streamREFUSED05 GB10 GB14 GB2 GB — the cap this used to be2h 1080p @10 Mbps ≈ 8.4 GB1h 4K @30 Mbps ≈ 12.6 GB1h 4K60 @50 Mbps≈ 22 GBBoth ceilings must hold. 2 GB spread across 3 hours is 1.6 Mbps — which is why theold 2 GB cap and the 3-hour claim beside it could not both be true.

Where 14 GB comes from. It is a measured number, not a round one. The machine that renders is sized by its memory — 32 GiB — and a job needs room for your source plus roughly 2.2 times it for everything written alongside, because a final export can come out at about the size of the source. 32 ÷ 2.2 is 14.5. Raising the published cap without raising that first would only move the refusal later, to after you had spent forty minutes uploading. It covers an hour of 4K at 30 Mbps, ten minutes of 4K at 195 Mbps, and three hours of 1080p. It does not cover an hour of 4K60 at 50 Mbps, which is about 22 GB — for that, transcode down first.

If you are here because something refused a 2 GB file: 2 GB is about a seventh of this ceiling, and worth understanding as a number, because it used to be ours. A 2 GB cap sitting next to a 3-hour claim is arithmetic that does not close — 2 GB across 3 hours is 1.6 Mbps, and even a single hour only fits under 4.7 Mbps. An ordinary 1-hour 1080p recording runs 8 to 20 Mbps, so the product was refusing exactly the files its own copy said it accepted. If a tool publishes both a size cap and a duration cap, divide one by the other before you trust either. Full per-asset limits are in the upload documentation.

What Everyone Else Publishes

These are the limits each vendor states in its own documentation or pricing page, read on 4 August 2026. Where a vendor does not publish a number, the cell says so rather than guessing — and “not published” is itself the useful finding, because it means you discover the ceiling by hitting it. Limits change; check the source links under the table before you rely on a row.

ToolMax file sizeMax durationFree tier
Valmera14 GB per main video3 hoursSame ceiling — free accounts get 50 one-time credits, not a smaller file
Descript1 GB Free · 10 GB Hobbyist · 20 GB Creator · 50 GB BusinessNot published60 minutes of media per month
Opus Clip10 GB per local upload10 hoursNot published in the cited docs
VEED1 GB per file on Free · no limit on paid (5 GB per project recommended)No per-file limit; 4–5 hours per project recommended1 GB per file (per file, not per project)
Kapwing250 MB Free · 6 GB paidVideos over 2 hours cannot be processed250 MB per file · 720p processing
Submagic3,000 MB2 min Starter · 5 min Professional · 30 min BusinessNot published
Vizard10 GB600 minutes (10 hours)Not published
CapCut (web)Not publishedNot publishedNot published
GlingNot publishedNot publishedNot published
Premiere · DaVinci Resolve · Final CutYour diskYour diskNot a hosted service

Sources. Descript pricing (per-plan file size, monthly media hours) · Opus Clip help — local file uploads (10 GB, 10 hours) and video length · VEED help — uploading files · Kapwing help — uploading and is Kapwing free · Submagic help — size and length limits · Vizard docs — minimum setup. CapCut's web editor and Gling publish no upload ceiling we could find; both were searched and neither states one in help documentation.

Two patterns worth naming. First, the duration caps cluster where the product's unit of work is — clip generators publish generous hour figures because they only ever hold a few minutes of it at a time, while caption-first tools like Submagic cap at minutes because a Short is the deliverable. Second, the free tier is usually where the real wall is: 250 MB a file, or 1 GB a file, or one hour of media a month. Valmera's free plan gets the same 14 GB and 3 hours as every paid plan — what runs out on free is credits, not your file. A comparison of the editing itself rather than the caps is on Valmera vs Descript and Valmera vs Opus Clip.

Why Long Recordings Are Genuinely Hard

A three-hour file is not a three-minute file with more of it. Five things break at length, and each one has a mechanical cause you can ask any tool about.

The transcript stops fitting

Three hours of speech is roughly thirty thousand words. Handed to a model whole, with per-word timings attached, that is the bulk of a context window spent before the edit is discussed. So the transcript cannot be a single blob the agent holds — it has to be something the agent queries. Valmera returns transcript text against a 48,000-character budget per call and word-level timings 400 words at a time, each with an explicit instruction to narrow the range and call again. That is why a three-hour file behaves like a three-minute one at the point of a request.

Memory over hours of material

An editor who forgets what happened in hour one cannot tell you that the story in hour two is a repeat. Nothing in a language model's context survives that scale by default, so the durable memory has to live outside it: the index and the edit decision list are both on disk, both readable at any moment, and the agent re-reads the current program map every turn rather than recalling it. It is the difference between remembering the video and being able to look it up.

The narrative thread

Mechanical cleanup scales linearly and editorial judgement does not. Cutting every silence in a three-hour recording is the same operation eighteen hundred times; deciding which twenty minutes of it are the episode is one decision that requires having understood all three hours. This is the part no tool does for you honestly. What an agent can do is give you the structure fast enough that the decision is cheap — search the transcript, read the shape, then cut by description.

Render time is real time

Every preview of a long program costs machine time proportional to its length, which is why previews render from a smaller proxy and only the final export touches your original file. If a tool renders the full-resolution deliverable on every iteration, long-form editing in it is a queue rather than a conversation. Ask what a preview is rendered from — it is the single most predictive question about whether a tool is usable at length.

Cost scales with duration, and it scales twice

Longer footage means a longer index to build, and a longer transcript to read on every turn afterwards. A tool priced per minute of source charges you for the first; a tool priced per turn charges you for the second. Valmera charges credits per agent turn, metered from what the turn actually consumed — and charges nothing at all for indexing, previews or exports, which are the parts that scale hardest with length.

The first of those is the one with a picture. The agent never re-watches a three-hour file — it reads an index built from it once, in windows, with every window bounded so that no single call can flood the turn.

How a three-hour file is read: indexed once, then queried in windowsA three-hour source file is indexed once into five artifacts: sentences with speaker labels, word-level timestamps, measured silences, shot boundaries and labeled frame tiles. Indexing is charged no credits. Each agent turn then reads bounded windows of that index rather than the file — transcript text against a 48,000-character budget, word timings 400 words per call with an instruction to narrow the range and call again, a transcript search that returns matching sentences, and frames rendered on demand. The source file is read again only at export.THE FILE · READ TWICE, EVERTHE INDEX · BUILT ONCE, 0 CREDITSONE TURN · BOUNDED READS3-HOUR SOURCEEXPORTup to 14 GBnever modified~30,000 spoken wordsreplays the decision listagainst the originalsentences + speaker labelsword-level timestampsmeasured silencesshot boundarieslabeled frame tileson disk · re-read every turnget_transcriptget_wordssearch_transcriptlook_at48,000 chars per call400 words per callmatching sentences onlyframes, rendered on demandonceat the endEvery bounded read ends by telling the agent how much was withheld and how to ask for the rest,so a long file degrades into more calls rather than into a truncated answer it cannot detect.Indexing, previews and exports are charged no credits — only agent turns are.

The reason that last line matters: a truncation the agent cannot see is how long-video editing quietly goes wrong. A model handed the first third of a transcript with no marker will confidently edit as though it read all of it. Every bounded read here ends with the count of what was withheld and the exact call to get it, which turns a silent failure into an extra tool call. It is not glamorous and it is most of what makes three hours behave like three minutes.

What Changes at One Hour and at Three

Under about 20 minutes, nothing about length is a factor. The whole transcript fits in one read, indexing is quick, and a preview renders while you are still typing the next request.

At an hour, three things start to show. The transcript no longer fits in a single call, so the agent works by search and by window — which means your requests get better results when they carry a landmark (“the bit where she talks about pricing”) than when they carry a guess at a timestamp. Preview renders take long enough to notice, so batching several changes into one request beats five small ones. And the file is now big enough that upload time, not edit time, is the longest wait of the session.

At two to three hours, the editorial problem overtakes the mechanical one. Removing dead air across three hours is still one request; deciding which twenty minutes are the episode is the actual work, and it is yours. This is where finding the structure first stops being good practice and becomes the only order that works. It is also where silence detection pays for itself most — an unrehearsed three-hour conversation carries a startling amount of nothing.

Past three hours, Valmera refuses, and says so before you upload. Split the recording, or transcode it and split it. Two 90-minute halves index independently and edit independently; what you lose is the ability to cut across the seam in one request, since the edit decision list addresses one main video.

How to Work a Long Recording Down

There is an order to this, and it is not the order most people try. Find the structure, cut the dead air, then the takes, then the tangents, and only then polish. Every pass makes the next one easier to judge, and doing them backwards — polishing a stringout, or captioning before cutting — is how a long edit becomes a re-edit.

An illustrative two-hour recording worked down pass by passAn illustrative worked example, not a measurement. A two-hour stringout is reduced by removing dead air to about one hour forty-four, then filler words to about one hour forty-one, then repeated takes to about one hour thirty-six, then one tangent to about one hour twelve. The final polish pass changes the look and sound but not the length.ILLUSTRATIVE · 2-HOUR CONVERSATION · RUNTIME AFTER EACH PASSstringout2:00:00− dead air1:44− filler words1:41− repeated takes1:36− one tangent1:12polish1:12length unchanged

PASS 1 — FIND THE STRUCTURE

Before removing anything, get the shape. Ask what the recording covers and where each topic starts; the agent reads the transcript and answers with real timestamps. This is the only pass that changes nothing, and skipping it is why long edits go wrong — you cannot decide what to cut until you know what is there.

PASS 2 — DEAD AIR

Silence removal is measured, not judged, so it is safe to run across the whole file in one request. On an unrehearsed two-hour conversation it typically removes more than any other pass. Cuts land on word boundaries, so nothing sounds clipped, and every cut is reversible.

PASS 3 — FILLER AND TAKES

Ums, uhs and false starts next, then repeated takes: where a line was said twice because the first attempt was flubbed, the transcript shows the phrase twice and the second one is usually the keeper. Both are still mechanical — the agent measures rather than decides.

PASS 4 — TANGENTS

Now the editorial cut, and the first one that needs you. Name the section to lose in plain language — “cut the ten minutes about the conference” — and the agent finds it in the transcript and removes it. Do this after the mechanical passes, because a tangent is easier to spot in a tight program than in a stringout full of pauses.

PASS 5 — POLISH

Loudness, music under the intro and outro, a grade, punch-ins on the moments that carry weight, and captions if the deliverable wants them. Polish last for a mechanical reason: music has to be fitted to a program length that does not exist until the cuts are done, and captions timed against the raw recording drift by the length of everything removed before them.

Passes 2 and 3 can be one request — silences, filler words and captions in one sentence is a single turn, and the agent sequences the dependency itself. Pass 4 cannot, because a tangent is a judgement and the agent has no way to measure one off the file. The individual operations have their own pages: removing silence, removing filler words and removing repeated takes.

How to Edit a 2-Hour Recording with AI

  1. 1
    Upload the whole file
    Up to 14 GB or 3 hours, in MP4, MOV, MKV or WebM. Do not pre-trim it — the agent cuts better with the whole recording indexed, and anything cut can be restored. Both ceilings are checked before the upload starts.
  2. 2
    Let it index once, and read the structure
    Indexing builds a word-level transcript with speaker labels, measured silences, shot boundaries and labeled frame tiles. It scales with duration, shows progress, costs no credits, and is reused by every later edit. Then ask what the recording covers and where each topic starts.
  3. 3
    Cut the mechanical passes in one request
    "Cut every silence longer than half a second and remove the filler words." On an unrehearsed two-hour conversation this is the single biggest reduction, and it is measured rather than judged, so it is safe to run across the whole file.
  4. 4
    Cut the tangents by describing them
    Name the section in plain language — "drop the ten minutes about the conference, it starts around 38 minutes" — and the agent finds it in the transcript and removes it. This is the pass that needs you, because a tangent cannot be measured off the file.
  5. 5
    Polish, then export from the original
    Loudness, music under the intro and outro, a grade, punch-ins. Polish last, because music has to be fitted to a program length that does not exist until the cuts are done. The export renders from your original upload at source quality and costs no credits.

Indexing is the only wait that scales with length, and it happens once. Every request after it reads the index rather than the video.

A Worked Plan: One 2-Hour Podcast

Two people, two hours, one camera, no script. Here is the whole session as requests. The runtimes are an illustration of how the passes stack rather than a measurement of your recording, and the turn count is what this sequence takes when it goes well — a correction after a preview is a normal extra turn, not a failure.

TurnWhat you typeWhat comes back
1“What does this cover, and where does each topic start?”Structure, with timestamps. Nothing edited yet.
2“Cut every silence longer than half a second, and the filler words.”≈ 2:00 → ≈ 1:41. One pass over the whole file.
3“Find the retaken lines and keep the last attempt of each.”≈ 1:41 → ≈ 1:36.
4“Drop the section about the conference — it starts around 38 minutes.”≈ 1:36 → ≈ 1:12. The first cut that needed you.
5“Tighten the open: start on the first real question.”Cold open. Usually one or two corrections after the preview.
6“Master to −14 LUFS, chill music under the intro and outro only.”Levels and bed. Ducked under speech.
7“Punch in on the three moments you think land hardest, and show me.”The agent proposes; you veto what you disagree with.
8Export.Rendered from the original 2-hour file at source quality. Not charged.

What it costs. Credits are charged per agent turn and metered from what that turn actually consumed — there is no fixed per-turn price, so quoting one here would be a made-up number. What is worth knowing is which parts are free and which scale. Indexing the 2-hour file: free. Every preview render: free. The final export: free. Only the eight turns above are charged, and a turn on a long recording costs more than the same turn on a short one because it reads more transcript and looks at more frames. The practical consequence is that a long-form month is priced by how many times you ask, not by how many hours you uploaded — which is the opposite of how the per-minute clip tools price. The pools and how they are spent are documented on the credits page, and the plans are on pricing.

Then the clips. The eight turns above produce the long cut. Shorts come out of the same project afterwards, and each one is its own request: “cut the exchange about pricing into a 9:16 aspect ratio Short with karaoke captions” is one turn plus a correction or two. Ten Shorts is therefore ten directed requests, not one batch — genuinely more work than an automated clip maker that ranks candidates for you, and genuinely more control over which moment gets cut and how.

Honest Limits at Length

The things that get harder, not easier, when the file is long — and the ones that do not work at all.

  • No multi-cam sync. A three-camera podcast has to arrive as one already-cut file. The edit decision list addresses one main video, not a bin of angles, so there is no sync-by-audio and no angle switching. This is the most common reason a long-form shoot does not fit here.
  • One deliverable per request. A 2-hour episode into ten Shorts is ten requests. No batch output and no ranked candidate list. If volume is the point, an automated clipper is the right shape.
  • Captions are burned in. No SRT or VTT import or export, and no chapter metadata — which for a 2-hour episode is usually exactly what you wanted, and we do not have it.
  • No audio repair. No denoise, no “studio sound”, no per-speaker leveling, and no separating music out of an already-baked track. Two hours of a bad room stays two hours of a bad room. Loudness mastering and per-layer gain exist; restoration does not.
  • Speaker labels, not diarization you can name. The transcript marks who is speaking as S0 and S1, which is what makes “keep only the guest's answers” work — but you cannot attach names to them or level them independently.
  • Past 3 hours it refuses. Not gracefully degrades — refuses, before the upload. Split the file. And 14 GB is a real wall too: an hour of 4K60 at 50 Mbps is about 22 GB and needs transcoding first.
  • Upload time is yours. A 10 GB file over a domestic connection is a long wait no product can shorten, and it happens before indexing starts.
  • English is the best-tested transcription path, and on a long recording the transcript is what the whole edit is built on.

The agent knows these edges as well: an out-of-scope request on a long file is refused rather than faked, by the same server-side check that stops it claiming an edit it did not make. The full toolset — 108 tools, 97 editing and 11 session — is on the tool reference, and the same registry is drivable from Claude over MCP if your long-form workflow already lives in a chat window.

Who Hits These Numbers

Podcasters recording two-hour conversations and cutting them to an episode plus clips — the podcast workflow covers speaker labels and retakes in more detail. Course creators with hour-long modules recorded in one take. Webinar and conference organisers turning a recorded session into something watchable. Streamers cutting a three-hour VOD down to the parts worth keeping, then to YouTube Shorts. Every one of them has the same shape of problem: the raw file is the asset, it is far too long, and most of the runtime is removable by measurement before any judgement is required.

Frequently Asked Questions

Judge it on three published numbers, not on marketing: the maximum file size, the maximum duration, and what a preview is rendered from. Most AI video editors are built for clips and cap uploads in minutes — Submagic allows 2 minutes on Starter and 30 minutes on its top tier, Kapwing refuses videos over 2 hours, and Descript scales the per-file size limit from 1 GB to 50 GB by plan. Valmera takes a single file up to 14 GB or 3 hours on every plan including free, indexes it once, and renders previews from a proxy so iteration stays fast at length. If your recording is longer than 3 hours, Opus Clip publishes a 10-hour ceiling and Vizard publishes 600 minutes, though both are clip generators rather than editors of the long cut.
Yes. A 2-hour recording is well inside Valmera's 3-hour ceiling, and at ordinary 1080p bitrates it is roughly 8–9 GB, inside the 14 GB ceiling too. The workflow is: upload, wait once for indexing, then edit by describing what you want. A realistic full pass on a 2-hour podcast — structure, silences, filler, repeated takes, one tangent removed, loudness, music, punch-ins, export — is about eight agent turns. What length changes is the indexing wait at the start, not the speed of each request afterwards.
3 hours per uploaded video. The limit is checked before you spend time uploading rather than after, because refusing a 5-hour file at the end of a 40-minute upload is not a kindness. Duration is the constraint that costs real machine time — it drives how long indexing takes and how long every render takes — so it is bounded on purpose and separately from file size.
14 GB for the main video, on every plan including free. Extra clips are capped at 500 MB, music at 50 MB and images at 10 MB. The 14 GB figure is measured rather than rounded: the render machine's working directory is bounded by its 32 GiB of memory, a job needs room for the source plus roughly 2.2 times it for what it writes alongside, and 32 ÷ 2.2 is 14.5. Raising the number without raising that first would only move the refusal to after you had finished uploading.
Yes — 2 GB is about a seventh of Valmera's 14 GB ceiling. If a tool is refusing your 2 GB file, the cap is theirs and it is usually a plan tier: VEED limits free accounts to 1 GB per file, Kapwing to 250 MB. Worth knowing where 2 GB comes from as a number, because Valmera's own cap used to be 2 GB and it was incoherent: 2 GB spread across 3 hours is 1.6 Mbps, and even a single hour only fits under 4.7 Mbps, so an ordinary 1-hour 1080p recording was being refused by a product whose own copy said it accepted 3-hour videos.
Because their unit of work is a clip. A tool that generates vertical shorts only ever needs to hold a few minutes at a time, so a short cap costs it nothing and saves it transcription, storage and render time. The cap becomes a wall the moment your source is the long recording itself — a webinar, a podcast, a lecture, a stream. That is a different job: the long file is the input, and everything short comes out of it afterwards.
Indexing scales with the length of the file and shows progress throughout — a 3-hour upload takes meaningfully longer to index than a 5-minute one. It happens once. Every edit afterwards reads the index rather than the video, so a request against a 3-hour file is not three hours of work, and it is not slower than the same request against a short one. Indexing is not charged in credits.
Per turn, somewhat — a turn on a long recording reads more transcript and often looks at more frames, and credits are metered from what the turn actually consumed rather than charged at a fixed price. But the three operations that scale hardest with duration — indexing, preview renders and the final export — cost no credits at all. In practice the cost of a long-form project is driven by how many turns you take, not by how many hours you uploaded.
Yes, and this is where an agent and an automated clipper genuinely differ. Valmera produces one deliverable per request, so ten Shorts is ten directed requests — “cut the exchange about pricing into a 9:16 Short with karaoke captions” — each aimed at a moment you chose. An automated clipper produces a batch of candidates in one pass and lets you curate. If you want volume, the clipper is the right shape; if you want a specific moment cut well, the agent is.
No multi-cam sync — a three-camera shoot has to arrive as one cut file, because the edit decision list addresses one main video rather than a bin of angles. No batch output, so a long recording into ten Shorts is ten requests. No SRT or VTT export and no chapter metadata; captions are burned into the picture, which for a 2-hour episode is usually not what you want. No team seats or share links. And the transcript is best tested in English.

Upload the Two-Hour File

14 GB or 3 hours on every plan, free included. Index it once, then edit by describing what you want.

Start free →
See pricing →

Related Articles

AI Video Editing for Podcasters
The long-form workflow end to end: speaker labels, dead air, retakes, and the clips that come out afterwards.
Auto-Edit a Podcast Video
Silences, filler words and captions from one sentence, sequenced so captions are timed against the edit rather than the raw recording.
Remove Silence From Video
The single pass that removes the most runtime from an unrehearsed recording, measured rather than judged.
Uploading Footage
Formats, the 14 GB / 3-hour ceiling, per-asset limits for clips, music and images, and importing from a link.
Agentic Video Editor
The mechanism underneath: the index, the edit decision list, and the loop where the agent looks at the frames it produced.