Design docs used to come after the code. Now they come before, and they're half the work. Marcus, an engineer at a fintech in Stockholm, hits this constantly: it's 11pm, he's in Cursor outlining a payment settlement redesign, voice flowing, ideas crystallizing on the screen. Then he hits Wispr's free cap, thousands of words in, mid-thought, and the thinking stops.
He switches to typing. The prose fragments. By morning, it needs rewriting.
This is not a speed problem. It's a flow problem.
The developer workflow shifted, but tools didn't follow
Code used to be the output. Now intent is. Prompts, specs, PR descriptions, incident postmortems, you're writing English for a model to interpret. Cursor doesn't make you type faster. It helps you write clearer specifications so the model understands better.
Voice is faster than keyboard for this. Not marginally. Enough to matter. Enough that losing the thread mid-sentence costs you hours of cleanup and rework the next day.
Why cloud dictation solved the wrong problem
Wispr, Willow, and Superwhisper all follow the same pattern: a nice interface over Whisper, priced at $8-14 monthly, with a hard ceiling on free usage.
Wispr: 5,000 words free, then upgrade. The others: similar caps, similar paywalls.
They optimized for speed when the real problem was flow. But there's another cost: your words leave your machine. Infrastructure details, security context, proprietary code snippets, none of that goes to someone else's server. Developers who care about this (which is most of them) just skip the tool entirely.
Local Whisper removes both constraints
Architecture matters here. Whisper-large-v3 achieves 96.3% word accuracy on LibriSpeech. When you run it on your device, there's no variable cost. No caps. No metering.
You hit 2,400 words, 10,000 words, 50,000 words, the cost stays zero because the computation happens on your machine.
This is not a limitation dressed up as a feature. This is a deliberate choice. Pro tier becomes the cloud rewrite layer, the feature that polishes rough speech into clean prose. But that's enhancement, not the base.
The base is: uninterrupted flow for as long as you need it.
Why Marcus switched to Cursor
He chose it specifically for autocomplete. Voice is faster to record than to revise, and Cursor's suggestions reduce rewrites significantly. He refuses cloud dictation entirely because explaining to security why proprietary code lives on Wispr's servers is not a conversation he wants to have.
If a tool respects that he owns his own words, stored locally, sent only where he decides, he'll use it for the work that matters. The design documents. The incident postmortems. The long-form thinking that can't be interrupted.
The honest trade-off
Local Whisper is not better at everything. Cloud rewrite is genuinely useful for polish. Paid services make sense in certain workflows. But the baseline should not be a cap. The baseline should be: as long as you're thinking, the tool records. As long as you're flowing, nothing interrupts.
That's what changes.