285 karma · joined January 24, 2025
Highly recommend their Edge Containers product, super simple and has nice primitives to deploy globally for a low latency workloads.
We connect all containers to one redis pubsub server to push important events like user billing overages, top-ups etc. Super simple, very fast, one config to manage all locations.
At one point, I carefully designed a spec document, forced Opus to reread it, create a plan with the planning tool that followed the spec, and use the task tool to track the implementation... AND AFTER OPUS READS THE FIRST FUCKING FILE, it says, "Oh, there are missing dependencies in project X. It’ll be hard to add them, so I’m going to throw away the whole plan and just do a simple fix..."
After that, I canceled my $200 Max plan, which I’d been subscribed to since June 2025, and decided to check out Codex
My favorites are Gemini 3 Flash and Mistral Voxtral Transcribe 2. Gemini when I need special formatting and clean-up, and Voxtral when I need fast input (mostly when working with AI).
Actually, I'm experimenting with this kind of stuff and trying to find a nice UX to make Ottex a voice command center - to trigger AI agents like Claude, open code to work on something, execute simple commands, etc.
I haven't found official benchmarks yet, but you can find Gemini 3 Flash word error rate benchmarks here: https://artificialanalysis.ai/speech-to-text/models/gemini — they are close to SOTA.
I speak daily in both English and Russian and have been using Gemini 3 Flash as my main transcription model for a few months. I haven't seen any model that provides better overall quality in terms of understanding, custom dictionary support, instruction following, and formatting. It's the best STT model in my experience. Gemini 3 Flash has somewhat uncomfortable latency though, and Flash Lite is much better in this regard.
We benchmarked it for real-life voice-to-text use cases:
<10s 10-30s 30s-1m 1-2m 2-3m
Flash 2548 2732 3177 4583 5961
Flash Lite 1390 1468 1772 2362 3499
Faster by 1.83x 1.86x 1.79x 1.94x 1.70x
(latency in ms, median over 5 runs per sample, non-streaming)
- 1.8x faster than Gemini 3 Flash on average
- ~1.4 sec transcription time for short to medium recordings
- ~$0.50/mo for heavy users (10h+ transcription)
- Close to SOTA audio understanding and formatting instruction following
- Multilingual: one model, 100+ languages
Gemini is slowly making $15/month voice apps obsolete.
It provides a significant quality increase over Gemini 2.0 Flash-Lite and Flash-Lite models, matching Gemini 2.5 Flash performance across key capability areas:
Improved response quality: Aims to match 2.5 Flash performance and align with target Flash-Lite use cases.
Improved instruction following: Targeted improvements to serve as a reliable migration path for complex chatbot and instruction-heavy workflows.
Improved audio input: Improved audio-input quality for tasks like Automated Speech Recognition (ASR).
Expanded thinking support: You can control how much reasoning the model performs by choosing from minimal, low, medium, or high thinking levels. This feature lets you balance response quality and speed for your specific use case.
---
Already available in Google AI Studio and OpenRouter
The top feature is the per-app custom settings - you can peak different models and instructions for different apps and websites.
- I use the Parakeet fast model when working with Claude Code (VS Code app). - And I use a smart one when I draft notes in Obsidian. I have a prompt to clean up my rambling and format the result with proper Markdown, very convenient.
One more cool thing is that it allows me to use LLMs with audio input modalities directly (not as text post-processing). e.g. It sends the audio to Gemini and prompts it to transcribe, format, etc., in one run. I find it a bit slow to work with CC, but it is the absolute best model in terms of accuracy, understanding, and formatting. It is the only model I trust to understand what I meant and produce the correct result, even when I use multiple languages, tech terms, etc.
I'm working on a Wispr/Spokenly competitor. It's free without any paywalled features, supports local models and a bunch of API providers including Mistral.
For local models ottex has - parakeet V3, Whisper, GLM-ASR nano, Qwen3-ASR (don't have voxtral yet though, looking into it).
btw, you can try new voxtral model via API (the model name to pick is `voxtral-mini-latest:transcribe`), I personally switched to it as my main default fast model - it's really good.
I'm working on a Wispr/Spokenly competitor. It's free without any paywalled features, supports local models and bunch of API providers including Mistral.
btw, the model name to pick is `voxtral-mini-latest:transcribe`, I personally switched to it as my main default fast model - it's really good.
I recently added support for Mistral provider, model is actually a very good one, I personally switched to it as my default model.
p.s. the app is free for personal user, has support for both local models and BYOK with OpenRouter, Groq, Mistral, Fireworks, and more coming soon.
Native macOS and iOS apps with OpenRouter BYOK. Same quality as in proprietary products for 1-3$ per month instead of 35$.
My experience also points to compiled languages that give immediate feedback on build. It’s nearly impossible to stop any AI agent from using 'as any' or 'as unknown as X'casts in TypeScript - LLMs will “fix” problems by sweeping them under the rug. The larger the codebase, the more review and supervision is required. TS codebase rots much faster then rust/C#/swift etc.
I love global voice-to-text transcription (especially when working with Claude Code or Cursor) and simple AI shortcuts like "Fix Grammar" and "Translate to {Language}".
I realized I was spending around €35/mo (€420 a year) on two apps for AI features that cost just pennies to run.
So I built Ottex - a native macOS app with a tiny footprint. Add your OpenRouter API key and get solid voice-to-text using Gemini 2.5 Flash, plus any OpenRouter model for AI shortcuts.
- Transcribe voice to text (especially useful when you need to explain something to Claude code )
- (soon) select text to instantly Check grammar / Improve writing / change tone of text
- (soon) select text to Translate between languages
I discovered that I have a few 10/20$ subscriptions (grammarly, raycast, wisperflow) that do embarrassingly simple stuff I can one shot with cheap SLM. So I decided to build a one app specialized in small repetitive tasks on computer.
https://techcrunch.com/2024/12/26/microsoft-and-openai-have-...
I inclined to trust the business which earns money from me - this means they are aligned with value I get and there is little incentive to break the trust and a high stakes to keep the trust when you get paid to be trustworthy.
I trust more the greedy capitalists than politicians in this question because I don't understand incentives of the latter. At least the business model is fairly transparent - you can check the company and how it makes money, in reverse incentives of the governments and their officials is broken - to get elected, get rich, get power, not lose job and keep producing new laws and regulations because if you want to keep your job you can't say “Everything is working, the best thing I can do right now is to monitor the system, collect the data and do nothing for a few years”.
e.g. 4k/sec saturates PG CPU to 95%, you get only 20% on redis at this point. Now you can compare latencies and throughput per $.
In the article PG latencies are misleading.
Comparing throttled pg vs non-throttled redis is not a benchmark.
Of course when pg is throttled you will see bad results and high latencies.
A correct performance benchmark would be to give all components unlimited resources and measure performance and how much they use without saturation. In this case, PG might use 3-4 CPUs and 8GB of RAM but have comparable latencies and throughput, which is the main idea behind the notion “pg for everything”.
In a real-world situation, when I see a problem with saturated CPU, I add one more CPU. For a service with 10k req/sec, it’s most likely a negligible price.
Really recommend to check them out if you need a vector DB. I tried qdrant and zilli cloud solutions and in terms of operational simplicity turbopuffer just killing it.