Wispr Flow review: pricing, free limits, privacy and Canto

This review is built from what could be checked on November 3, 2026: Wispr's own site, pricing page, help center, terms, privacy pages and changelog; its Canto research post; reporting by TechCrunch and Outlook Business; reviews by Cult of Mac, Android Police and How-To Geek; benchmarks published by two competitors, Superwhisper and Willow; user reports on Hacker News and Google Play; an independent forensic write-up of the Mac app; and the alternatives' own pricing pages. Each piece of evidence says who produced it, and the numbers Wispr gives about itself are labelled as its own.

What Wispr Flow is

Wispr Flow is a dictation app from Wispr AI, Inc., a San Francisco company led by CEO Tanay Kothari. You hold a shortcut (Fn on a Mac, Ctrl+Win on Windows), speak, let go, and Flow pastes finished text into whatever field has the cursor. On iPhone it is a keyboard; on Android it is a floating bubble that sits beside your normal keyboard. Every dictation is transcribed on Wispr's servers, so it needs a connection on every platform (help center).

The main draw is the cleanup. Flow drops "um" and "uh", applies spoken corrections ("4 pm, actually 3 pm"), turns spoken items into lists, and writes in a Formal, Casual or Very casual style depending on whether you are in email, work chat or personal messaging. A personal dictionary and voice snippets handle names and boilerplate. Wispr's homepage claims Flow is four times faster than typing: 220 words a minute against 45.

Since August 2026 the same app also includes Notetaker, which records meetings from your computer's audio instead of sending a bot into the call and writes summaries you can query (TechCrunch) or hand to Claude and ChatGPT over MCP (Wispr). If meetings are your main need, compare it with the tools in our Otter.ai alternatives list.

What changed in 2026

Wispr had a busy and uneven year:

• February 23: Flow launched on Android, after what Wispr calls a waitlist of hundreds of thousands, with a model for Hinglish (TechCrunch, Wispr).
• March: Wispr dropped Delve, the compliance platform behind its earlier SOC 2 work, after Delve was accused of producing fake compliance evidence (TechCrunch); more on that below.
• Late May to early July: outages and an accuracy drop, which Wispr's changelog acknowledged in five notes between June 4 and July 9.
• August 5: Notetaker launched on Mac; Windows followed on September 15.
• August 17: Wispr raised $280 million in a Series B at a $2 billion valuation, led by Menlo Ventures, bringing its total to $361 million (TechCrunch). Outlook Business, citing Reuters, reported the same terms and a valuation nearly three times the $700 million of nine months earlier (Outlook Business).
• September 17: Flow switched every user to Canto, Wispr's own speech model.
• October 2: model training was switched off and locked for Growth and Enterprise teams.

The spring problems matter because they explain the Canto launch. Wispr's changelog blamed early-June outages on "a combination of vendor failures and our own scaling issues", said an Auto Cleanup setting had been too aggressive for some users, and said some dictations had been handled by older models, one of them too aggressive about changing words. When the funding was announced, TechCrunch wrote that several users had complained about a quality dip in the preceding weeks. On Hacker News, a developer who works on a competing app posted a re-test of 525 fixed clips: Flow's word error rate rose from 9.0% in May to 11.2% in June (Hacker News).

Canto: Wispr's own speech model

Wispr built Canto in its new research lab, led by chief scientist Ariya Rastrow, and says it trained the model for real conditions: laptop microphones, earbuds, background voices and whispering. Its research post compares it with other speech models on Wispr's own data (Wispr AI Lab):

All figures are Wispr's. The test set is English only, comes from users who opted into data sharing, and the post, read November 3, 2026, does not publish it, so others cannot rerun it; even its size differs between the research post (10 hours) and the changelog (9.8 hours). On public read-speech sets, Wispr says Canto tied for best on LibriSpeech and did not lead on FLEURS or Common Voice.

Canto handles speech recognition; rewriting your words is a separate step. Baseten, which sells Wispr inference, says in an undated case study that Flow used fine-tuned Llama models for that cleanup, with a target under 700 milliseconds end to end (Baseten). Wispr's public pages, checked November 3, 2026, name no other AI provider and point to the subprocessor list in its Trust Center.

Plans and what the free tier allows

From the pricing page:

The pricing page describes the free tier only as "Limited words per week". The numbers are in the help center: on Mac and Windows a warning appears after 2,000 words and new dictations are blocked above 5,000; Android slows after 2,500 and stops at 3,000; iPhone slows after 550 and stops at 750. The week resets on a fixed day for everyone, unused words do not carry over, and Wispr says the allowances can differ by account. Command Mode, the voice-editing feature, needs a paid plan or trial (help center).

New accounts get a 14-day Pro trial (30 days by referral, 90 for students). Students with an .edu email get three months of Pro, then pay $6 a month billed annually (Wispr). On cancellation, Pro runs to the end of the period and the help center says there is no automatic prorated refund; the terms updated August 19, 2026 say refunds are issued only if required by law (terms).

What public tests and users report

• Cult of Mac, January 31, 2026. After about a month, Rajesh Pandey reported accurate results through a low-gain microphone more than an arm's length away, with music playing, and when whispering in a cafe (Cult of Mac).
• Android Police, February 25, 2026. Sanuj Bhatia wrote more than 3,000 words on Android in about a week and called it the best voice typing he had used across devices; he flagged the need for a connection (Android Police).
• How-To Geek, April 26, 2026. In a seven-app Windows comparison, Jorge A. Aguilar measured about 800 MB of RAM at idle and an 8 to 10 second start, and saw punctuation and grammar errors in Microsoft Teams (How-To Geek). Wispr argues that cloud processing spares the battery that on-device models would drain (Wispr).
• Superwhisper's benchmark, a competitor's test. With Flow's cleanup off, runs in August 2026 averaged 8.7% word error rate and 1.28 seconds to text over eight datasets, against 6.6% for Superwhisper's own S1 Voice and 7.8% for OpenAI GPT Transcribe (Superwhisper). The runs predate Canto.
• Willow's benchmark, a competitor's test. Willow says Flow took about 400 milliseconds on short dictations and 900 on average-length ones in September 2026, against 190 to 230 for Willow (Willow).
• Store ratings. On November 3, 2026 the iPhone app had 4.8 stars from about 18,700 ratings (Apple App Store data) and the Android app 4.6 from 14.5K reviews with more than a million downloads (Google Play). A September 2026 Play review praises the PC app but says Android sometimes leaves text on the clipboard instead of inserting it, a fallback Wispr's help center also describes.
• User comparisons. On the Hacker News thread for Canto, one user who had moved through four dictation apps called Flow the least accurate of them (Hacker News).

Privacy: what leaves your device

Wispr documents its privacy controls in detail, but the defaults favour data collection. From its security overview, data-sharing article and Data Controls page:

• Training. "Improve the model for everyone", formerly called Privacy Mode, lets Wispr use your audio, transcripts and edits to train its models. It is on by default for Free and Pro and can be switched off; it is off and locked on Growth and Enterprise. Wispr says its third-party AI providers may not train on your data.
• Storage. Dictation cloud storage is a separate switch. With it off, Wispr says audio is discarded once transcribed (privacy page). Desktop history can be kept, deleted after 24 hours or never saved.
• Metadata. Even with training off, Flow uploads word count, duration, app, URL, timestamp, microphone and edit statistics for each dictation.
• Screen context. Context Awareness, on by default, reads text around the cursor, on-screen text and the app or site you are in. Wispr's help center says it works from text and does not take screenshots; with it off, the desktop app still sends the app name and website domain.
• Where data goes. Wispr says data is processed and stored in the United States on AWS and that it offers no offline or on-premises version (enterprise privacy page). The same page says that on iPhone, "background Flow" keeps the microphone recording between dictations so you do not have to switch apps, and transcribes only when you press the button.

Two of Wispr's own statements do not line up: Data Controls says Wispr has zero data retention agreements with all third-party AI providers for dictation, while the enterprise privacy page says processors generally delete data within 30 days.

An independent forensic write-up of the Mac app, version 1.4.752, published in April 2026, matches Wispr's description more than it contradicts it: logs of the apps and websites in use, on-screen accessibility text sent with the audio, hourly metadata uploads with sharing off, and a local database holding 198 MB of audio (forensic write-up). Its database had a screenshot field that the author found unfilled. The same write-up traced 145 swallowed spacebar presses in ten minutes to Flow's keyboard hook, which watches for the dictation shortcut system-wide.

Security claims and the Delve episode

Wispr now says its SOC 2 Type II report and ISO 27001 certificate come from A-LIGN and that HIPAA-ready dictation is available on any plan with a click-through Business Associate Agreement (privacy page). Per its help center, Wispr does not countersign the BAA, and signing it locks training and cloud storage off and blocks Notetaker (help center).

These audits are new. Wispr's March 27, 2026 post says it moved from Delve to the Drata platform and hired A-LIGN to re-verify its SOC 2 and HIPAA status (Wispr). The allegations TechCrunch reported said Delve's customers were routed through audit firms that rubber-stamped reports Delve had generated; Delve disputed them (TechCrunch, March 22, 2026). Wispr's help center also says Notetaker launched after the audit periods and was not assessed, while the privacy page says the report covers every Wispr product.

For developers

On Mac and Windows, Flow recognises variable and function names visible in Cursor, Windsurf and VS Code, and turns spoken file names into @-tags in Cursor and Windsurf chat. Per the help center, variable recognition needs the IDE's screen-reader mode on, and file tagging does not work in VS Code or Copilot chat (help center). On Mac, long dictations into Claude Code and Codex in the terminal are pasted as plain text. Command Mode rewrites selected text or drafts replies by voice, and sends the text box contents to Wispr with the instruction.

Who should not use Wispr Flow

• Anyone who needs dictation offline, on an air-gapped machine or with data kept outside the United States.
• Lawyers, clinicians and others with strict data rules who cannot change the defaults or sign a BAA, and anyone who needs Notetaker under HIPAA.
• People who dictate mostly on an iPhone and do not want to pay: the free cap stops at 750 words a week.
• Owners of iPads, Linux PCs, Chromebooks or Windows on ARM laptops, which Wispr's system requirements exclude.
• Buyers who want a one-time price or a guaranteed refund.

Alternatives and when they fit better

Prices from each product's own page, checked November 3, 2026:

Wispr Flow's case against these is breadth: the same cleanup, dictionary and styles on Mac, Windows, iPhone and Android, plus coding features and a meeting notetaker in one subscription. If you need privacy by default, offline use or a one-time price, one of the alternatives above fits better.