Developers now type prompts and specifications instead of code, longer thinking on the same keyboard. Most voice tools hit a word cap exactly when that thinking gets deepest. The interruption costs more than you'd expect.
A developer in Stockholm discovers the paywall
Marcus, a backend engineer at a Series B fintech, drafted an incident postmortem at 11pm last month. Forty minutes of voice notes, speaking faster than he could type, capturing the exact sequence of what broke and why. Three thousand two hundred words, raw and unpolished, but complete.
Then his voice tool showed a paywall: word cap reached.
The flow stopped. The thinking stopped. He rewrote parts the next morning from memory. Some nuance was gone. This is what broken free-tier economics looks like: it catches you at the exact moment you're most productive.
The bottleneck moved
For decades, voice software marketed itself as a typing speed hack. Speak faster than you type. Get words on screen quicker. That framing worked when the work was voice-to-document, period.
The work is different now. If you're coding with Claude, Cursor, or Copilot, you're no longer shipping code from voice dictation. You're writing specifications, design decisions, and intent. You're explaining what you want the AI to build. You're commenting code in retrospect. You're drafting PR descriptions and Slack threads that explain what you changed and why.
This work is inherently longer. A two-minute conversation about a settlement race condition becomes a five-minute voice memo, becomes a three-thousand-word design doc. A bug investigation that takes twenty minutes to explain takes the full twenty minutes to explain well; compressing it into a hundred typed words loses the narrative.
The old metric, words per minute spoken, stopped being the real constraint. The constraint is finishing the thought. And almost every cloud-based voice tool in the market cuts you off before the thought is complete.
Why word caps break momentum
Wispr Flow ($14/month) caps its free tier at two thousand words per month. Willow ($12/month) caps at three thousand. Superwhisper ($8.49, indie) caps at four thousand. These aren't typos. Four thousand words per month is roughly one hundred words per day, or a single four-minute voice session.
The economics are intentional. Cloud transcription costs variable compute. Every word transcribed costs the company money. So free tiers are metered. The paywall exists at the technical boundary, not the user's boundary.
Here's what happens in practice: You start a voice session at 10:45pm. You're in Cursor, explaining the payment flow while your hands design the schema. Fifteen minutes in, you're two thousand words deep, and the microphone stops listening. The cap is reached. You switch to typing. By the time you've typed fifty words, the thinking has shifted. The continuity breaks. What was natural becomes fragmented.
This happens at the exact wrong moment: when you're deepest in the problem.
Local changes the economics
Recitey runs Whisper locally on your Windows device. It's the same model OpenAI open-sourced, and it hits 96.3% word accuracy on the LibriSpeech benchmark. No cloud upload. No variable cost per word.
This means no word cap. No metering. No paywall hiding in the free tier.
Marcus refuses cloud transcription for payment settlement code. Spending years thinking about data classification makes you uncomfortable sending banking code intent to a cloud API. Local transcription isn't a feature preference; it's a compliance checkpoint. And it turns out that preference aligns exactly with what works technically: if the transcription happens on your machine, the marginal cost of ten thousand words versus one thousand is essentially zero.
This is why the pricing model is different. The expensive part isn't transcription. It's what comes next: polishing the raw voice output into publishable prose. The rough voice draft gets sent to a model that rewrites it in under two seconds. That's the value that costs compute. That's what lives in Pro.
Free tier is the dictation machine. Pro is the rewrite engine.
What you actually get
With no word cap, Marcus can voice a full incident postmortem in one session. Eight thousand words, raw, unfiltered, still in voice-first grammar. The output lands in Slack, his text editor, a browser form, anywhere he can paste because Recitey works through the Windows clipboard. It works in Cursor, Linear, Notion, any tool that accepts text.
The trade-off is that you're getting a first draft. Cloud dictation tries to normalize your voice as it transcribes. Recitey doesn't. What you get is what you said: accurate, but rough. Filler words. Thinking-out-loud grammar. Repeated phrases.
Some people hate this. They want clean-sounding output. They want the tool to do free work it can't actually do for free.
Marcus doesn't care. He's not shipping voice drafts. He's shipping explainers and specs he'll read back, adjust, and send. He wants the bottleneck to be his thinking, not his free tier limit. He wants to speak a three-thousand-word design doc without an interruption. And he wants the company's payment code to stay on his machine.
For that workflow, thinking out loud and capturing intent with no word limits, local-first, the structure is different. Free is genuinely usable, and Pro adds something you might want later, not something you need to escape the free tier.