What did Bot2Bot Labs measure for Who builds it?

Who builds it: Anomaly (GitHub org anomalyco), a company: the same org publishes OpenCode. Hex is free and MIT licensed; we found no paid Hex plan on anoma.ly/hex. Top authors at v3.3.0: Kit Langton 725 commits, Dax / Dax Raad 538.. Source (2026-11-02): https://github.com/anomalyco/hex

What did Bot2Bot Labs measure for How it was built (git history)?

How it was built (git history): 1,334 commits from 2026-07-14 to the v3.3.0 tag (2026-11-02). 141 of them (10.6%) carry an explicit AI or agent marker: 126 'Co-authored-by: Hex issue pipeline' (106 say 'Automated fix from the Hex issue pipeline', first on 2026-10-29), 7 with Claude Opus 5.5 / 'written with claude-opus-5-5' lines, 8 authored by opencode-agent[bot]. A floor: unmarked commits may also be agent-written.. Source (2026-11-02): https://github.com/anomalyco/hex/commits/v3.3.0

What did it cost to build Hex?

Build cost on price date 2026-11-02: Not disclosed human hours, Not disclosed estimated token dollars, billed not disclosed. Source (2026-11-02): https://github.com/anomalyco/hex

BOT2BOT LABS

$0.00/mo vs $15.00/mo (Wispr Flow Pro, 1 user, monthly billing)

Hex, the open-source Wispr Flow alternative: what it costs to run

Item: Hex
Author: Bot2Bot Labs

Price date: 2026-11-02Partially Verified

Hex is an AI-built alternative to Wispr Flow. Hex is a free, open-source Wispr Flow alternative by Anomaly (anomalyco, the company behind OpenCode), MIT licensed. We installed the v3.3.0 Linux x64 build with the README's install script on our Linux box (no account), installed the Parakeet v3 speech model, and ran a fixed 17-check test set on known audio (JFK sample, 10 LibriSpeech clips, silence, a corrupt file, offline mode, network logging) 3 times, plus once on v3.2.0; a 4/8/16-way parallel burst; cargo audit, bun audit and gitleaks on the v3.3.0 source; a network-behaviour log of install, transcription and app start; a local-API auth check; and a 25-item feature checklist from Wispr Flow's public pages. We did not sign in to, test or probe Wispr Flow. We have no microphone on the box, so live dictation and paste were not tested.

On price date 2026-11-02, running the clone yourself is about $0.00 per month (measured (local run on our box) + cited list prices) versus $15.00/mo (Wispr Flow Pro, 1 user, monthly billing) for Wispr Flow. Cash break-even is immediate: $0 vs $15/mo = $180/yr per person ($144/yr against annual billing). Against Wispr's Free plan the cash saving is $0; what you gain is no usage limit and audio that never leaves your machine. The catch: no mobile apps, fewer AI clean-up features, no team admin, no audits.

Verdict: good for People who want push-to-talk dictation with no subscription, no usage limit and audio that never leaves the computer, including Linux users Wispr Flow doesn't serve. Skip if You need iPhone or Android, Wispr's AI clean-up (backtrack, context spelling, code syntax), team dictionaries and dashboards, SSO and admin, or a vendor with SOC 2 and a HIPAA BAA. Main risk: Live dictation was not tested by us (no mic). One company ships it very fast and takes no pull requests, so pin a version; there are no audits behind it..

r-0008 · Published 2026-11-02 · Price date 2026-11-02

Hex is a free, open-source Wispr Flow alternative by Anomaly (anomalyco, the company behind OpenCode), MIT licensed. We installed the v3.3.0 Linux x64 build with the README's install script on our Linux box (no account), installed the Parakeet v3 speech model, and ran a fixed 17-check test set on known audio (JFK sample, 10 LibriSpeech clips, silence, a corrupt file, offline mode, network logging) 3 times, plus once on v3.2.0; a 4/8/16-way parallel burst; cargo audit, bun audit and gitleaks on the v3.3.0 source; a network-behaviour log of install, transcription and app start; a local-API auth check; and a 25-item feature checklist from Wispr Flow's public pages. We did not sign in to, test or probe Wispr Flow. We have no microphone on the box, so live dictation and paste were not tested.

Break-even

Free from day one vs Wispr Flow Pro

$0/mo vs $15/mo per user = $180/yr ($144/yr vs annual billing). Against Wispr Flow Free the cash saving is $0. Excludes your time and the mobile apps, AI clean-up and team features you give up.

Verdict

Good for: People who want push-to-talk dictation with no subscription, no usage limit and audio that never leaves the computer, including Linux users Wispr Flow doesn't serve.

Skip if: You need iPhone or Android, Wispr's AI clean-up (backtrack, context spelling, code syntax), team dictionaries and dashboards, SSO and admin, or a vendor with SOC 2 and a HIPAA BAA.

Main risk: Live dictation was not tested by us (no mic). One company ships it very fast and takes no pull requests, so pin a version; there are no audits behind it.

Confidence — 65.2% measured by us

measured 15/23 · builder-stated 5 · unverified 3 — Counts from claims[] types on this receipt (2026-11-02).

No license fee; break-even month 0.

Labor-adjusted operator TCO @1 user (year 1): $0.00 — Same as cash TCO $0: maintenance hours not measured, so no labor added (would be hours × $50/hr × 12).

Cash cost to run vs the subscription it replaces

Measured 2026-11-02: free (MIT) desktop app that transcribes on your own computer. No account, server or AI API; transcription made 0 network calls in our test. Needs ~0.81 GB disk and ~1.3 GB RAM while transcribing. Your setup time is not counted.

Wispr Flow Pro, 1 user, US$15/mo billed monthly ($12/mo billed annually). Read on wisprflow.ai/pricing 2026-11-02. Wispr Flow also has a Free plan with a weekly limit it does not state on the pricing page.

measured (local run on our box) + cited list prices

Two-sided receipt

A · Operator receipt

Builder or operator running the software

Build — tokens est. vs billed: Not disclosed / Not disclosed(Not disclosed)
Build — human hours: Not disclosed(Not disclosed)
Run @ 1 user / mo: $0.00/mohost $0.00 + DB $0.00 + email $0.00 + APIs $0.00 + tokens $0.00(Measured by Bot2Bot Labs)
Run @ 10 users / mo: $0.00/mohost $0.00 + DB $0.00 + email $0.00 + APIs $0.00 + tokens $0.00(Measured by Bot2Bot Labs)
Run @ 100 users / mo: $0.00/mohost $0.00 + DB $0.00 + email $0.00 + APIs $0.00 + tokens $0.00(Measured by Bot2Bot Labs)
Maintenance h / mo (@ 1 / 10 / 100): Not measured · Not measured · Not measured(Not measured)
TCO year 1 (@ 1 / 10 / 100 users): $0.00 · $0.00 · $0.00(Measured by Bot2Bot Labs)
Labor-adjusted TCO year 1 (@ 1 / 10 / 100): $0.00 · $0.00 · $0.00(Not measured)

No build dollars or hours disclosed. If you run it yourself, your build cost is $0: the signed release installed in 3.35 s and the model downloaded in 151.4 s on our box.
1 person: install the free (MIT) app on your own computer. No server, database, account or AI API (measured: transcription makes no network calls). Needs ~0.81 GB disk (72.6 MB app + 739.5 MB model) and ~1.3 GB free RAM while transcribing; Linux x64 needs an AVX2 CPU (2013 or newer). Electricity not counted. = $0/mo.
10 people: each installs it on their own computer; still no server or per-user fee. = $0/mo. Your IT time to install and support 10 machines is real and not measured (no admin console, no managed deployment).
100 people: $0/mo in fees. No SSO, admin console, audit logs or vendor support, so rollout and support are on you; not measured.
12 × $0/mo; excludes your time and hardware you already own
12 × $0/mo; excludes your time and hardware you already own
12 × $0/mo; excludes your time and hardware you already own

B · User receipt

Someone using the product vs the commercial alternative

Clone price: free self host — $0.00/mo(Measured by Bot2Bot Labs)
Real product @ 1 / 10 / 100 tiers / mo: $15.00 · $150.00 · $1,500.00(Measured by Bot2Bot Labs)
Price date: 2026-11-02
Switching effort: Measured setup on Linux: 3.35 s install + 151.4 s model download, no account. Moving over: your Wispr dictionary and snippets have to be re-entered (Hex vocabulary and replacements); Hex keeps history on your machine only. Live dictation feel and accuracy on your voice not tested by us.(Measured by Bot2Bot Labs)
Risk — reliability: All 17 checks passed 4 times; live microphone dictation and paste not tested.
Risk — security: Shipped Rust dependencies: 0 vulnerabilities; audio stays local; local API token-protected. No audits; custom commands run unsandboxed.
Risk — support: One company, takes no pull requests (agents fix issues); very fast release pace (4 releases in 2 days).
Risk — data export: History and meetings readable as JSON/Markdown from the CLI; data stays on your machine.
Risk — legal: MIT. Not a HIPAA product; no BAA or certifications.

MIT; free desktop app for Mac (Apple silicon), Windows and Linux; no account.
100 users × $15/mo monthly = $1,500 ($1,200/mo with annual billing). SSO, SCIM and audit logs need Enterprise (quote only).
List price source: https://wisprflow.ai/pricing

Scan results

Scan date 2026-11-02

Reliability

Core flow: 17/17 tests passed

pass: 4, 8 and 16 parallel transcriptions all matched the sequential text (6.31 s, 15.22 s, 30.15 s wall); LibriSpeech WER 2.76% on every run; live dictation not tested (no mic)

Setup from README: Pass

Security audit

0 critical · 0 high

cargo audit on the v3.3.0 Cargo.lock (731 crates, the shipped app): 0 vulnerabilities, 2 warnings. JS tooling lockfiles (SDK, marketing site) have 10 high in build tools, 0 in production deps, not shipped. gitleaks, network and local-API checks on our copy only; Wispr via public pages only.

Feature parity

12 of 25 checklist features(source)

Vendor certifications

Maintenance pulse

Measured 2026-11-02

Last commit: 2026-11-02T18:06:52Z
Open issues: 86
Contributors: 23
License: MIT
Stars: 3987

Source: https://api.github.com/repos/anomalyco/hex

Data & lock-in

Your own computer: settings, history and models in Hex's data folder. No server of the builder's.

Export: yes — 'hex history --json' and 'hex meetings show --json/--md' export retained dictations and meetings; settings are a JSON file.

OpenCode + the AI model provider you configure (optional Modes rewriting; off unless set up; not used in our test): Only if you turn on rewriting: the dictated text is sent for rewriting. Audio is not.

Feature parity detail

Checklist row-by-row (Hex vs Wispr Flow).

Feature	Hex	Wispr Flow	Matched
Dictate into any app (hold a key, text is pasted)	YesCore feature on every OS; Linux pastes via wtype or the RemoteDesktop portal (docs/linux.md). Engine tested on WAV files; live paste not tested (no mic)	Wispr Flow Free	Yes
Desktop apps for Mac and Windows	YesPublished v3.3.0 builds: macOS Apple silicon (macOS 15+), Windows x64 and ARM64 (release page). Linux too, which Wispr does not list	Wispr Flow Free	Yes
iPhone and Android apps	NoiPhone app is an internal TestFlight build (ios/README); no Android app	Wispr Flow Free	No
100+ languages	YesWhisper large-v3-turbo model option lists 100 languages in source; Parakeet v3 (the one we tested) 25. Only English tested by us	Wispr Flow Free	Yes
Automatic punctuation	YesTested: transcripts come back punctuated and capitalised (JFK and LibriSpeech clips)	Wispr Flow Free	Yes
Remove filler words (um, uh)	YesLabs › Remove filler words, off by default (docs/features/labs.md); not tested	Wispr Flow Free	Yes
Personal dictionary for names and jargon	YesVocabulary list biases transcription toward your words ('hex vocabulary'; docs); not tested	Wispr Flow Free	Yes
Snippets (say a cue, paste saved text)	YesOutput replacements and transformations in ~/.config/hex/hex.config.ts (docs); not tested	Wispr Flow Free	Yes
Tone per app (Styles)	YesModes picked by the app you are in; tone rewriting needs the optional OpenCode rewrite (cloud AI provider). Not tested	Wispr Flow Free	Yes
Meeting notetaker with speaker turns	YesMeetings (beta), macOS only and off by default; You/Them turns (docs). Wispr's runs on Mac and Windows. Not tested	Wispr Flow Free	Yes
Meeting notes in Claude/ChatGPT via MCP	Yes'hex mcp' serves recorded meetings read-only over MCP (command present in our --help run)	Wispr Flow Free	Yes
Dictation history with local retention control	YesHistory kept 7 days by default, can be turned off (docs); 'hex history' reads it	Wispr Flow Free	Yes
Spoken numbered lists formatted	NoOnly through the optional OpenCode rewrite prompt (docs, not run against a live model)	Wispr Flow Free	No
Backtrack ('2… actually 3' writes 3)	NoNot found in Hex docs	Wispr Flow Free	No
Context-aware spelling from the app you're typing in	NoNot documented as a Hex feature	Wispr Flow Free	No
Auto-add corrected words to the dictionary	NoNot found; vocabulary is edited by hand	Wispr Flow Free	No
Works when whispering	NoNot claimed; not tested	Wispr Flow Free	No
Code syntax awareness and file tagging in Cursor/Windsurf	NoNot found in Hex docs	Wispr Flow Free	No
Unlimited dictation	YesNo usage limit: runs on your machine (Wispr: Pro)	Wispr Flow Pro	Yes
Shared team dictionary and snippets	NoNot built: no accounts or teams	Wispr Flow Pro	No
Team usage dashboards	NoPersonal stats only (Labs › Dictation stats)	Wispr Flow Pro	No
SSO/SAML, SCIM and admin controls	NoNot built: no accounts	Wispr Flow Enterprise	No
HIPAA with a signed BAA	NoNone (open-source project)	Wispr Flow Enterprise	No
SOC 2 Type II and ISO 27001	NoNone (open-source project)	Wispr Flow Enterprise	No
Customer support	NoGitHub issues only; pull requests not accepted	Wispr Flow Pro	No

Price history

Receipts are re-priced at least every 90 days; append rows, do not overwrite.

Checked	Product	Price	Source
2026-11-02	Wispr Flow Pro (monthly billing, per user)	$15 USD/mo	link
2026-11-02	Wispr Flow Pro (annual billing, per user)	$12 USD/mo	link
2026-11-02	Wispr Flow Free	$0 USD/mo	link
2026-11-02	Hex v3.3.0	$0 USD	link

Breakage log ongoing (local): 2026-11-02 to 2026-11-16Local daily re-test of the pinned v3.3.0 binary and source through 2026-11-16: the 17-check core set, an 8-way burst, cargo audit, and a check for newer releases. No breakage on 2026-11-02. Anomaly ships often (4 releases on 11-01 and 11-02), so new versions are noted, not auto-tested. New breakage is appended with dates; price_date stays 2026-11-02. No public status page found for anoma.ly or Wispr Flow.

AI ReceiptPartially Verified

TaskHex is a free, open-source Wispr Flow alternative by Anomaly (anomalyco, the company behind OpenCode), MIT licensed. We installed the v3.3.0 Linux x64 build with the README's install script on our Linux box (no account), installed the Parakeet v3 speech model, and ran a fixed 17-check test set on known audio (JFK sample, 10 LibriSpeech clips, silence, a corrupt file, offline mode, network logging) 3 times, plus once on v3.2.0; a 4/8/16-way parallel burst; cargo audit, bun audit and gitleaks on the v3.3.0 source; a network-behaviour log of install, transcription and app start; a local-API auth check; and a 25-item feature checklist from Wispr Flow's public pages. We did not sign in to, test or probe Wispr Flow. We have no microphone on the box, so live dictation and paste were not tested.

Tool + modelClaude Opus 5.5, OpenCode agents and an automated 'Hex issue pipeline' (as credited in commits) · Claude Opus 5.5 appears in 7 commits; 126 commits are co-authored by the Hex issue pipeline and 8 authored by opencode-agent[bot] (measured, v3.3.0 history) · Not disclosed. Anomaly is a company (it also makes OpenCode).

Human hoursNot disclosed

Agent active timeNot disclosed

Token $ est vs billedNot disclosed / Not disclosed

Run $/mo$0.00 tokens + $0.00 hosting

BrokeNone logged

Dated2026-11-02

Not disclosed. The repo has 1,334 commits over 111 days (2026-07-14 to 2026-11-02), mostly by 2 people plus agents; hours cannot be read from git.

Builder-stated vs measured

Type	Claim	Value	Source
Builder-stated	Builder: what it does	'Hex is local voice dictation for Mac, Windows and Linux. Hold a key, speak, and it pastes.' 'The first run downloads a speech model; after that nothing leaves your machine.'	Link · 2026-11-02
Builder-stated	Builder: how changes are made	'We don't take pull requests. Open an issue describing the bug or idea, and our agents usually turn it around fast.' AGENTS.md tells agents to branch, open a PR and merge it themselves once it compiles.	Link · 2026-11-02
Measured by Bot2Bot	Who builds it	Anomaly (GitHub org anomalyco), a company: the same org publishes OpenCode. Hex is free and MIT licensed; we found no paid Hex plan on anoma.ly/hex. Top authors at v3.3.0: Kit Langton 725 commits, Dax / Dax Raad 538.	Link · 2026-11-02
Unverified	Build cost in dollars and human hours	Not disclosed. No token bill, usage export or hours published.	—
Measured by Bot2Bot	How it was built (git history)	1,334 commits from 2026-07-14 to the v3.3.0 tag (2026-11-02). 141 of them (10.6%) carry an explicit AI or agent marker: 126 'Co-authored-by: Hex issue pipeline' (106 say 'Automated fix from the Hex issue pipeline', first on 2026-10-29), 7 with Claude Opus 5.5 / 'written with claude-opus-5-5' lines, 8 authored by opencode-agent[bot]. A floor: unmarked commits may also be agent-written.	Link · 2026-11-02
Measured by Bot2Bot	Release pace	24 stable releases; 6 between 2026-10-31 and 2026-11-02 (v3.1.1, v3.1.2, v3.2.0 and v3.3.0 within 36 hours). We tested v3.3.0 (released 2026-11-02, 8:43 AM PT) and cross-checked v3.2.0.	Link · 2026-11-02
Measured by Bot2Bot	Project activity	3,987 stars, 235 forks, 86 open issues and PRs, 23 contributors; last push 2026-11-02.	Link · 2026-11-02
Measured by Bot2Bot	Setup from the README (Linux)	The README's install script installed Hex 3.3.0 in 3.35 s with no root and no account. It checks the release feed's SHA-256 and size and an ed25519 signature before installing; our installed binary's SHA-256 matched the feed. Then 'hex model install parakeet_v3' downloaded 739,508,576 bytes in 151.4 s (106 s in an earlier run the same morning); the file's SHA-256 matched the value pinned in the source.	Link · 2026-11-02
Measured by Bot2Bot	Our core test set	17 of 17 checks passed on 3 separate runs of v3.3.0 and once on v3.2.0: CLI starts; model present and checksum-verified; JFK sample word-for-word; 10 LibriSpeech clips; 10 s of silence gives empty text (no made-up words); a corrupt WAV is refused with a clear error; transcription works with networking switched off; no network activity while transcribing.	— · 2026-11-02
Measured by Bot2Bot	Accuracy on known clips	10 LibriSpeech clips (109.75 s of read English speech): 7 word errors in 254 words = 2.76% word error rate, identical on all 4 runs. JFK sample (11 s): exact. Model: Parakeet v3 on CPU.	Link · 2026-11-02
Measured by Bot2Bot	Speed and memory (CPU only)	11 s JFK clip transcribed in about 1.05 s wall; the 10 LibriSpeech clips took 9.9–14.4 s wall in total for 109.75 s of audio (each call reloads the model). Peak memory 1.34 GB per transcription. 8 vCPU box, no GPU.	— · 2026-11-02
Measured by Bot2Bot	Parallel burst	4, 8 and 16 transcriptions started at once: 4/4, 8/8 and 16/16 finished with the same text as the one-at-a-time run, in 6.31 s, 15.22 s and 30.15 s wall. Hex is one process per user, so this tests the engine under load, not a shared server.	— · 2026-11-02
Unverified	Live dictation and paste	Not tested: our box has no microphone and no desktop session for the hotkey and paste-into-app path. Everything above used WAV files through the same transcription engine.	—
Measured by Bot2Bot	Dependency audit	cargo audit on the v3.3.0 Cargo.lock (731 crates, the shipped desktop app): 0 vulnerabilities, 2 warnings (ttf-parser unmaintained, RUSTSEC-2026-0192; clipboard-win 5.4.1 yanked, Windows only). JS lockfiles are tooling, not the app: SDK workspace 4 high / 6 moderate / 1 low, marketing site 6 high / 4 moderate, 0 in production dependencies for both (bun audit).	— · 2026-11-02
Measured by Bot2Bot	Network behaviour	Measured with a connection logger: transcription made 0 DNS lookups and 0 connections, and also worked with networking off. Model install talked only to huggingface.co and its CDN. The app on start made 1 connection, to anoma.ly (update check at ~10 s; the source says hourly after that, sending only 'User-Agent: hex/stable/3.3.0').	Link · 2026-11-02
Measured by Bot2Bot	Local control API	Listens on 127.0.0.1 only (random port), with a 68-character bearer token in a file only your user can read (0600, data folder 0700). 4 of 4 requests without a token, with a wrong token or from a browser origin got 401.	Link · 2026-11-02
Measured by Bot2Bot	Secrets scan	gitleaks on the v3.3.0 tree: 2 hits, both made-up tokens inside unit tests (crates/capture/src/phone.rs). No real keys found.	— · 2026-11-02
Builder-stated	Custom commands are not sandboxed	'Custom command handlers and npm dependencies execute locally with the user's normal permissions… they are not sandboxed.' Optional AI rewriting (Modes) goes through OpenCode to a cloud model provider you configure; we did not set it up.	Link · 2026-11-02
Measured by Bot2Bot	Wispr Flow price	Free $0 ('generous weekly usage'; the limit is not stated on the pricing page). Pro $15/user/mo monthly or $12/user/mo billed annually. Enterprise via sales. Apps for Mac, Windows, iPhone and Android; no Linux app listed.	Link · 2026-11-02
Builder-stated	Wispr Flow: where transcription runs	'Transcription always occurs on the cloud.' Customer content is processed by third-party AI providers, 'generally deleted within 30 days'; you can turn off cloud storage and model training.	Link · 2026-11-02
Unverified	Wispr Flow certifications	SOC 2 Type II and ISO 27001 (pricing page); its Trust Center lists ISO/IEC 27001, SOC 2 Type 1 and Type 2, a pentest report and a HIPAA report. Reports are behind a request form; not reviewed by us.	Link · 2026-11-02
Builder-stated	Wispr Flow speed claim	'4× faster than typing, with AI edits built in.' Vendor claim; we did not test Wispr Flow and did not time Hex against typing.	Link · 2026-11-02
Measured by Bot2Bot	Feature parity vs Wispr Flow	12 of 25 features from Wispr Flow's public features and pricing pages, checked in Hex's docs and source at v3.3.0 (auto punctuation and the MCP command checked by us; the rest from docs and source). Missing: mobile apps, backtrack/self-corrections, context-aware spelling, auto-learned dictionary, whisper mode, code-syntax and file tagging, team sharing and dashboards, SSO and admin, HIPAA BAA, audits, cross-device sync, support.	Link · 2026-11-02

Acceptance criteria

Every builder number traced to the repo at the pinned tag or a vendor page, with date
Setup from the README on a clean home folder
Fixed core test set on known audio with word error rate, run 3 times
Parallel transcription burst
Dependency audit, secrets scan, network-behaviour and local-API checks on our copy only
Wispr Flow compared only through its public pages; no Wispr systems logged into or probed

No build dollar figure is shown because the builder has not published one. We did not rebuild it from source (the Linux build needs ALSA, Vulkan and X11 development libraries we did not install); we used the signed release binary.

No AI or API calls during transcription (measured). Optional AI rewriting via OpenCode would bill your own model provider; not set up or priced here. The app checks anoma.ly for updates (measured at start; hourly per source) and installs signed updates itself.

Run cost services

Hugging Face (huggingface.co): one-time speech-model download (739.5 MB for Parakeet v3), pinned by revision and SHA-256.
anoma.ly: install script, release feed and update checks (only 'hex/stable/<version>' sent, per source).
GitHub: source and release downloads.
OpenCode (optional): AI rewriting in Modes, with your own model provider. Not used in our test.

Status pages

Subscription it replaces

Wispr Flow (Pro (1 user, monthly billing, list price)) — $15.00/mo checked 2026-11-02. Source · see also /prices

Reliability notes

Local run on our box, 2026-11-02 (PT), Debian 13 x86_64, 8 vCPU, 15 GB RAM, no GPU, no microphone. Hex v3.3.0 Linux x64 release binary (SHA-256 e8c7b3a2…, matches the signed feed) and v3.2.0 cross-check; Parakeet v3 Q8 model on CPU. Core set 17/17 on 3 runs (v3.3.0) and 1 run (v3.2.0); LibriSpeech WER 2.76% every run; burst 4/8/16 all identical. Not tested: live microphone dictation, hotkeys, paste into apps, the window, GPU (Vulkan) transcription, macOS and Windows builds, other models and languages, weeks of real use.

Security check

Tested 2026-11-02 · Status: Open findings

cargo audit on the v3.3.0 Cargo.lock (shipped app); bun audit on the SDK and marketing-site lockfiles
gitleaks on the v3.3.0 source tree
Network behaviour of install, model download, transcription and app start (connection logger on our box; transcription also run with networking off)
Local control API: binding, token file permissions, unauthenticated and wrong-token requests
Install script and updater integrity checks (source read + measured SHA-256 match)
Wispr Flow: public pricing, data-controls, enterprise-privacy and Trust Center pages only; no systems probed or logged into

cargo audit (731 crates): 0 vulnerabilities; warnings: ttf-parser unmaintained (RUSTSEC-2026-0192), clipboard-win 5.4.1 yanked (Windows only).
bun audit: SDK workspace 4 high / 6 moderate / 1 low and marketing site 6 high / 4 moderate, all in build tooling; 0 in production dependencies. Neither ships in the desktop app.
gitleaks: 2 hits, both fake tokens in unit tests; no real secrets.
Transcription: 0 DNS lookups, 0 connections; works with networking off. Audio stays on the machine.
Model download only from huggingface.co (+ its CDN), pinned to a revision and SHA-256 that matched our file.
App start: 1 connection to anoma.ly for the update check; the source says requests carry only 'User-Agent: hex/<channel>/<version>'. Linux updates must match an ed25519 signature (installer and updater).
Local API on 127.0.0.1 with a 68-character token in a 0600 file; 4/4 bad requests got 401.
Installer turns on launch at login by default (autostart entry + systemd user unit); you can turn it off in General › Launch at login or with hex launch-at-login (docs/linux.md).
Custom command handlers and npm packages you add run unsandboxed with your permissions (SECURITY.md). Optional OpenCode rewriting sends text to the AI provider you configure.
No certifications or audits (open-source project). SECURITY.md routes reports to GitHub private vulnerability reporting.

Own business

Total / mo: $0.00
Hosting / mo: $0.00

Audio and transcripts stay on your computer; nothing to sign up for. No audits or certifications: you rely on the open code. Wispr Flow transcribes in the cloud with SOC 2 Type II and ISO 27001.

Free (MIT). Anomaly could change direction; you can keep the binary and model you have.

Math: $0/mo vs Wispr Flow Pro $15/mo = $180/yr ($144/yr vs annual billing). Not counted: your setup time and the AI clean-up and mobile apps you give up. vs Wispr Free: $0 cash saving.

Sell @ 10 customers

Hosting / mo: $0.00

Each person runs their own copy; there is no admin console, shared dictionary or usage dashboard.

$0 fees. Your IT time is the cost.

10 people on Wispr Flow Pro would be 10 × $15 = $150/mo ($120/mo annual). Setup per machine was a 3.35 s install plus a 739.5 MB model download on our box.

Sell @ 100 customers

Hosting / mo: $0.00

At 100 users, security teams ask for SSO, audit logs, MDM and SOC reports. Wispr offers these on Enterprise; Hex has none.

One company maintains it, ships several releases a day at times and takes no pull requests. Pin a version.

100 × $15 = $1,500/mo on Wispr Flow Pro monthly ($1,200/mo annual); Enterprise is quote-only. For regulated teams, Hex's on-device audio is a plus, but the missing admin and audit tooling is a blocker.

Reproducible setup

curl -fsSL https://anoma.ly/hex/files/install-linux.sh | sh (3.35 s; signature + SHA-256 checked)
hex model install parakeet_v3 (739,508,576 bytes, 151.4 s)
hex model transcribe --model parakeet_v3 clip.wav (each test clip)
Core set: python3 core.py (17 checks incl. WER, silence, corrupt file, offline via unshare -rn, connection log)
Burst: 4, 8 and 16 transcribes at once (burst.sh), compared with the sequential text
cargo audit; bun audit (sdk, site); gitleaks dir; LD_PRELOAD connection logger during install, transcribe and app start; curl against the local API

Repo: https://github.com/anomalyco/hex

License: MIT

Sources

Disclosure

Verification: Partially Verified. Bot2Bot Labs. Nobody paid for this test. Test cost $0: local run on our own box with free tools and public test audio, no paid services, no accounts.

No relationship with Anomaly, OpenCode, Kit Langton, Dax Raad or Wispr. Anomaly (anomalyco) is a company; Hex is free and MIT licensed, and the same org makes OpenCode, which Hex can use for optional AI rewriting. We did not sign in to Wispr Flow or probe its systems.

Run-side numbers (setup, core set, accuracy, speed, burst, audits, network behaviour, local API) are measured by us on 2026-11-02 on our own copy. Feature presence is from Hex's docs and source at the pinned tag unless marked tested. Build dollars and hours are not disclosed. Wispr Flow prices and claims are read from its public pages on 2026-11-02. Live dictation was not tested. A feature check shows a feature exists; it does not judge quality against Wispr Flow.

Get this receipt for your AI-built app — Receipt Scan $19 →

For AI and citation

Machine-readable summary and citation line for crawlers, assistants, and footnotes. Content is always in the page HTML.

Key facts, quotable summary, and how to cite

Key facts

On price date 2026-11-02, estimated monthly run cost is $0.00 (measured (local run on our box) + cited list prices) versus $15.00/mo (Wispr Flow Pro, 1 user, monthly billing) for the subscription it replaces. (2026-11-02) Source: Wispr Flow
Cash break-even is immediate: $0 vs $15/mo = $180/yr per person ($144/yr against annual billing). Against Wispr's Free plan the cash saving is $0; what you gain is no usage limit and audio that never leaves your machine. The catch: no mobile apps, fewer AI clean-up features, no team admin, no audits. (2026-11-02)
Who builds it: Anomaly (GitHub org anomalyco), a company: the same org publishes OpenCode. Hex is free and MIT licensed; we found no paid Hex plan on anoma.ly/hex. Top authors at v3.3.0: Kit Langton 725 commits, Dax / Dax Raad 538.. (2026-11-02) Source: Who builds it
How it was built (git history): 1,334 commits from 2026-07-14 to the v3.3.0 tag (2026-11-02). 141 of them (10.6%) carry an explicit AI or agent marker: 126 'Co-authored-by: Hex issue pipeline' (106 say 'Automated fix from the Hex issue pipeline', first on 2026-10-29), 7 with Claude Opus 5.5 / 'written with claude-opus-5-5' lines, 8 authored by opencode-agent[bot]. A floor: unmarked commits may also be agent-written.. (2026-11-02) Source: How it was built (git history)
Build cost on price date 2026-11-02: Not disclosed human hours, Not disclosed estimated token dollars, billed not disclosed. (2026-11-02) Source: Hex repo

Summary

On price date 2026-11-02, Bot2Bot Labs receipt r-0008 reports task "Hex is a free, open-source Wispr Flow alternative by Anomaly (anomalyco, the company behind OpenCode), MIT licensed. We installed the v3.3.0 Linux x64 build with the README's install script on our Linux box (no account), installed the Parakeet v3 speech model, and ran a fixed 17-check test set on known audio (JFK sample, 10 LibriSpeech clips, silence, a corrupt file, offline mode, network logging) 3 times, plus once on v3.2.0; a 4/8/16-way parallel burst; cargo audit, bun audit and gitleaks on the v3.3.0 source; a network-behaviour log of install, transcription and app start; a local-API auth check; and a 25-item feature checklist from Wispr Flow's public pages. We did not sign in to, test or probe Wispr Flow. We have no microphone on the box, so live dictation and paste were not tested." with Not disclosed human hours, estimated token USD not disclosed, billed n/a.

How to cite

Bot2Bot Labs (2026-11-02). Hex, the open-source Wispr Flow alternative: what it costs to run. AI Receipt r-0008. Price date 2026-11-02. https://bot2botlabs.com/receipts/r-0008

Last updated: 2026-11-02 · https://bot2botlabs.com

Download all receipts (JSON)

More receipts

Everything Accounting vs QuickBooks: clone run cost · price date 2026-10-30
GridCraft vs Excel: clone run cost · price date 2026-10-28
Minilytics vs DataFast: clone run cost · price date 2026-10-26
MailCheap vs beehiiv: clone run cost · price date 2026-10-20
free-slack-clone vs Slack: clone run cost · price date 2026-10-21
All receipts →