AI tool comparison
Stet vs Sup AI
Which one should you ship with? Here is the side-by-side panel verdict, pricing read, reviewer split, and community vote comparison.
Productivity
Stet
Local macOS dictation that sounds like you — not like generic AI prose
75%
Panel ship
—
Community
Free
Entry
Stet is an open-source macOS dictation app that transcribes speech locally and then uses AI to clean up the output while actively preserving your personal writing style and tone. The core innovation is a voice model — a lightweight profile that learns from your past writing so the AI corrections don't flatten your voice into generic AI-ese. The result is meant to sound like you dictated it, not like it was passed through a generic LLM. The technical approach combines local Whisper-based transcription (nothing leaves your device during speech-to-text) with an optional AI refinement pass that can use your own API key (BYOK) or a $6.99/month subscription. The open-source release includes the voice profiling code, making it auditable and forkable. It's a direct response to Wispr Flow, which is closed-source and subscription-only. For writers, podcasters, and productivity users who dictate significant amounts of content, the voice preservation angle is genuinely differentiated. The proliferation of AI writing tools has created a recognizable 'AI voice' — flat, over-structured, and devoid of personality — that sophisticated readers are increasingly adept at detecting. Stet's bet is that preserving your actual voice is the most valuable thing an AI writing assistant can do.
AI Productivity
Sup AI
Runs 339 LLMs in parallel and downweights the hallucinating ones.
50%
Panel ship
—
Community
Free
Entry
Sup AI is an ensemble AI assistant that runs your query through 339 language models simultaneously, measures per-segment confidence across all responses, and synthesizes a final answer that amplifies agreement and suppresses likely hallucinations. The team claims a 52.15% score on Humanity's Last Exam (HLE) — 7.41 percentage points above the single best model — which, if verified, would make it the highest-scoring system on the benchmark to date. The underlying mechanism works like an LLM panel: each model votes on sub-claims within the response, confidence is estimated by agreement density, and the final output surfaces high-confidence segments while flagging uncertain ones. It's designed to reduce hallucination rate on factual tasks, not improve reasoning per se — the models in the ensemble aren't doing collaborative chain-of-thought, they're voting on outputs. Sup AI was built by Ken Mueller (Stanford, CEO) and Scott Mueller (AI Research Scientist) and launched on Product Hunt today. Pricing starts with $10 in free credits, no auto-charge, with a credit card required to start. The HLE benchmark claim is the headline and will face scrutiny — if verified, this is a meaningful research result. If it's cherry-picked, it's still a usable product with a differentiated architecture.
Reviewer scorecard
“Open-source, local-first transcription with BYOK is the right architecture. I've been burned by voice tools that upload my audio to servers I can't audit. The voice profile approach for preserving style is technically interesting — I want to see how it handles domain-specific jargon and code-switching between formal and casual registers.”
“The HLE claim needs independent verification, but the underlying ensemble approach is architecturally sound for factual Q&A tasks. Running 339 models is expensive — pricing will be the gating factor for production use. The $10 free credit is a fair trial.”
“The 'sounds like you' promise needs a lot of data to actually deliver — your voice profile is only as good as the writing samples it's trained on, and most people don't have a consistent, large corpus of their own writing. For casual dictators, this might just be Whisper with extra steps. Apple's built-in dictation is free and surprisingly good now.”
“Extraordinary claims require extraordinary evidence. A 7.41 point jump on HLE via ensembling — without publishing methodology — smells like benchmark gaming. The latency of running 339 models in parallel is also a real concern for anything other than async research tasks.”
“Voice-first computing is coming back, and the arms race for authentic AI writing assistance is heating up. The distinguishing factor won't be transcription accuracy — everyone has solved that — it will be voice fidelity. Stet is building in the right direction: local processing plus personal style models. Expect this architecture to be standard in two years.”
“Model ensembling is an underexplored direction in the race to reduce hallucination. If Sup AI's approach scales, it could be more durable than fine-tuning individual models — you get the wisdom of the crowd across model families, training data, and architectures simultaneously.”
“This is genuinely exciting for writers and content creators. The homogenization of AI-assisted writing is a real aesthetic problem — everything starts sounding like the same LinkedIn post. A tool that actively fights that tendency by learning your specific voice is solving the right problem. Even if the voice model needs work, the direction is exactly right.”
“For creative work, ensemble outputs tend to regress toward the mean — you get the most-agreed-upon version of something, which is usually the least interesting version. This is a tool for factual accuracy, not creativity. I'd stick with a single strong model for writing.”
Weekly AI Tool Verdicts
Get the next comparison in your inbox
New AI tools ship daily. We compare them before you waste an afternoon.