A dropped connection shouldn’t restart a big upload

A recording is almost uploaded. The connection drops. Starting again would send the whole file twice, even though most of it already arrived. Smol keeps a checkpoint of completed parts so you can reselect the original and continue.

That is the idea behind the homepage’s “≈10× less to resend”: if 90% of a file has been acknowledged, about 10% remains. The estimate describes retry bytes. It does not promise that your connection becomes ten times faster.

Keep the parts that arrived

Smol’s browser uploader splits a file into 10 MiB parts. The server stores them in a multipart upload: each part can arrive separately, and the storage service assembles the original when all parts are present. The browser saves the upload ID, completed part identifiers, and a hash of each completed part in a local checkpoint.

Consider a 100-part upload that stops after part 90. A full restart sends 100 parts again. With its checkpoint intact, Smol checks the original file and sends parts 91–100. A part that arrived without its response being saved may need to repeat; the useful unit of progress is an acknowledged part.

The browser uploader implements this loop. Progress survives an interruption through the saved checkpoint, rather than depending on a tab staying open forever.

Reselecting the file is a check, too

After a reload, the browser needs access to your local file again. Choosing one with a matching name and size is insufficient: its contents could have changed.

Smol uses a sampled fingerprint to find the upload session, then hashes every completed part of the reselected file before skipping it. A changed part stops recovery and asks for the original. This still reads local bytes, so resume has a validation cost even when it avoids sending those bytes across the network.

The server also checks the current user’s workspace and folder access on upload requests. A saved checkpoint does not grant permission after access is revoked. The isolated transfer results cover upload recovery across a Worker restart and original-file access after revocation. The per-part file check described here comes from inspection of the current browser uploader.

What the size numbers mean

On October 5, a live 101 MiB original download matched its SHA-256 hash after upload. The current upload route accepts files up to 25 GiB. That configured ceiling is about 250 times the live check’s size; it is not a completed 25 GiB benchmark.

Historical tests of Smol’s predecessor reached 12.25 GiB, but used a different upload path. They help explain the direction of the work, not the throughput of today’s app. The sanitized transfer data records the historical sizes, resumed throughput, current configured ceiling, and calculated retry scenario separately.

For a large recording, the practical benefit is fewer bytes to send after a failure. Keep the original available, reselect it when asked, and let Smol check which parts can be reused. Clearing browser storage removes the local checkpoint and its record of completed parts.

Comparing Google Drive and Dropbox

The homepage compares Smol with Google Drive and Dropbox. The predecessor tests above are engineering history, not a competitor benchmark. We checked the following official documentation on October 5, 2026.

Big uploads are already supported

Google Workspace documents a 5 TB per-file ceiling and a 750 GB upload/copy allowance per user within 24 hours. The first file that crosses that allowance can finish. Available storage still matters.

Dropbox’s upload guide lists 350 GB for web uploads and 2 TB for desktop uploads. Its separate size-limit guide warns that browser uploads above 375 GB may time out or be interrupted. These pages disagree on the web threshold, so we show both rather than silently treating 375 GB as a guaranteed browser limit.

Smol’s current ceiling is 25 GiB. Google Drive and Dropbox support much larger originals. Our 101 MiB live receipt does not establish performance at Smol’s ceiling or relative throughput against either competitor. The homepage chart converts published byte quantities to a common linear scale; it compares upload ceilings, not speed, preview size or transcription capacity.

Recovery is not unique to Smol

The Google Drive API supports resumable uploads. Dropbox upload sessions also upload files in parts and retain a session for up to seven days. These API capabilities do not prove how either product’s browser interface recovers after a reload.

Smol exposes recovery in its browser uploader through reselecting the original and checking saved parts. The roughly tenfold reduction in retry bytes at 90% completion compares resuming with a full restart. It is not evidence that Google Drive or Dropbox always restarts, or that Smol transfers ten times faster.

A two-hour interview exposes a different limit

Google Drive has searchable video transcripts for captioned videos. This is available to personal and Workspace accounts. Dropbox supports audio/video transcripts on dropbox.com, but documents a 105-minute duration ceiling. A two-hour recording is 120 minutes and exceeds that ceiling.

Dropbox Replay transcription is a separate feature of the Replay Add-On; its documentation also specifies 105 minutes and deletion 45 days after last access. We do not apply Replay’s paid access or retention policy to ordinary Dropbox transcription.

Smol’s current native media processing streams audio/video from storage, accepts media up to six hours, and transcribes five-minute chunks under a daily processing budget. The generic 2 GiB processing cap applies to non-media sources; audio and video bypass that byte cap. The media upload ceiling remains 25 GiB. These are configured ceilings, not completed six-hour or 25 GiB transcription benchmarks. Upload capacity, preview generation and transcription capacity must be compared separately.

Comments are common; the project conversation is the distinction

Google Drive supports file and timestamped video comments; it also documents a 1080p maximum playback resolution while retaining higher-resolution originals. Dropbox provides file comments and Replay media review. Neither should be presented as lacking feedback tools.

Smol combines ongoing folder chat with file comments, searchable processed text, and passage references tied to immutable file versions. Our comparison is about that combined workflow. It is not an assertion that transcription, media previews or collaboration are exclusive to Smol.

The real work behind this direction involves multi-gigabyte MOV/WAV recordings, thumbnails and ZIP handoffs. Dropbox already offers Extract All for archives, so archive extraction alone is not a unique Smol feature either. In Smol, completed extracted files remain usable while remaining entries process. Our large-archive recovery work is still being exercised on real files; it should not be advertised as a proven maximum-size benchmark.

Storage at cost and the team-plan comparison

Our pricing direction is to pass through Cloudflare R2 storage and transfer costs without markup. This is a pricing plan, not a claim that usage billing is already implemented. AI processing, application compute, request allocation and final billing terms are separate.

Cloudflare’s Standard R2 rates are $0.015 per GB-month stored, $4.50 per million Class A operations, and $0.36 per million Class B operations. R2 egress to the Internet is free, including through the Workers API. No per-GB egress charge does not mean every request or every application operation is free.

The account-wide Standard free tier includes 10 GB-month, one million Class A operations and ten million Class B operations per month. It is not a separate allowance for every Smol workspace. Cloudflare also rounds billable usage up to its billing units. The examples below use published storage rates before free allowances, request fees, rounding or tax.

Total physical R2 storage held for a month	R2 storage component
100 GB	$1.50
1,000 GB	$15.00
2,000 GB	$30.00

Physical storage includes retained originals, backups, historical versions and generated previews. If 100 GB of originals has a full 100 GB backup, those two copies alone total 200 GB and about $3/month at list storage rates. Storage pricing is not based only on the visible current-file sizes.

Three teammates with a 100 GB library

For a team that only needs a modest library, compare actual usage with the cost of purchasing seats:

Model	Monthly amount	What it includes
Smol’s planned R2 pass-through	Approximately $1.50 storage component	100 GB physical storage; requests, AI and compute separate
Google Workspace Business Standard	$42 subscription	Three $14 seats; 6 TB pooled allowance plus the Workspace suite
Dropbox Standard	$45 subscription	Three $15 seats; 5 TB pooled allowance plus Dropbox’s team features

Google’s annual-commitment price is $14 per user per month for Business Standard; its pricing page lists 2 TB pooled storage per user. Dropbox lists Standard at $15 per user per month on yearly billing. Its current Standard plan starts with 3 TB for the first member and adds 1 TB per additional actively-used license, with no minimum member count. The older three-seat minimum belongs to Standard Classic, not today’s new Standard plan.

These figures compare a storage component with full subscriptions. They do not establish that Smol’s complete bill is 28 or 30 times cheaper. Currency is USD, excluding taxes and introductory promotions; prices were checked October 5, 2026.

At high storage utilization the bundles can be cheaper: one Workspace Standard seat includes 2 TB for $14/month, whereas 2,000 GB in R2 storage alone is about $30/month. Dropbox Standard’s one-seat 3 TB allowance also has a lower effective storage rate when full. Teams already paying for Workspace may have no incremental subscription cost for their existing Drive allowance. The case for usage pricing is avoiding purchased capacity and seat-driven storage charges when your actual library is small.

Transfer cost and transfer limits are different

Repeated delivery from R2 adds no per-GB R2 egress fee. Reads still count as Class B operations; uploads and multipart requests count toward Class A operations. Thumbnails, previews, metadata and AI also use resources outside that storage calculation.

Dropbox’s shared-link and file-request bandwidth policy lists 1 TB/day for Standard and 4 TB/day for Advanced and Enterprise. Exceeding it can temporarily pause sharing. Those are service limits, not per-GB egress invoices. Google Workspace’s 750 GB daily upload/copy allowance discussed above is a different kind of limit; it should not be presented as a download bandwidth cap.