I've spent some time with Fable, and it is really good, definitely a step change from Opus 4.8, both for coding and general chat-style discussions. The vibes are incredible. There is an ease with which it solves problems and I've tested by replicating older chats in Fable - things that the older models found after 5-6 turns, Fable surfaces in the first response. It just gets things.
Apart from all the above: the fact that they are intentionally writing this (that they degrade frontier LLM dev, silently vs loudly for biology/cybersecurity) in the system card is interesting to say the least - especially just before IPO.
Notice that with this statement - that they're going to intentionally hobble the model for frontier LLM development - the general discussion has moved from, “Is the model actually that good?” to "they’re pulling the ladder up from behind them"
That's actually super smart - wonder if Mythos (or the next unreleased model) had a say in coming up with that strategy (if it's intentional). Also - having access to extremely capable models before anyone else - which they have by default - is a incredibly advantageous position to be in.
What we found was that for super fast tap to speak and paste text, WhisperKit is already close to instant (basically realtime for Apple Silicon). Faster than realtime is mostly only useful for batch processing of audio which is not really our product.
Some one asked this but got flagged so still answering it here:
ejoso 2 hours ago: The latency breakdown is honest and the right frame. Most of the delay budget gets eaten before and after the model, and that is harder to fix than it looks. Cold starts on local Whisper variants and WebSocket warmup on cloud paths are both worse than benchmarks suggest.
The differentiation question I keep landing on: whisper.cpp behind a hotkey with a paste shim is a solved afternoon project for a certain kind of person. The real gap is consistency across arbitrary apps, and that is genuinely harder than it sounds.
For the fully local path (flying, privacy-sensitive) what does Shoute add beyond well-packaged WhisperKit with better insertion handling? That answer is either the core pitch or an honest scoping of who this is for.
This makes a good point - latency and consistency are indeed the hardest to get right.
For the fully local path - the major value add for Shoute is still the speed and consistency, getting accessibility settings right and the flow from the model to text is useful for many folks. That and consistently keeping it updated with model updates as models improve and with OS updates.
I've been working on Shoute, a speak-to-text app for Mac and Windows that's built around one idea: the full loop has to feel instant
I do know this isn’t a new category. A lot of people here already have some version of this: whisper.cpp behind a hotkey, macOS dictation, SuperWhisper, Wispr Flow, or some other hand-rolled version.
I built one anyway because I kept bouncing off dictation tools in my actual workday.
My problem was not “can an app transcribe my voice?” Most of them can, and impressively well. The problem was the full loop:
press shortcut -> speak -> release -> cleaned up text appears where I was already typing - and that this happens consistently, quickly, day after day.
If that loop has enough delay, I lose the thread. If the output is too raw, I am back to editing. If the app needs screenshots to understand context, I start feeling uneasy about using it everywhere. You want to be confident that it always will work - or else you lose trust in it.
So the version I wanted was pretty narrow:
- it should feel super quick for short everyday dictation
- the output should be cleaned up before insertion
- it should work across ALL the usual apps
- it should never lose data
- it should support both local and cloud modes (personally for flying but privacy too for specific things)
- it should use only minimal context
Shoute solves all of that really well and is lightweight (native code) and fluid to use day to day. It has a generous free tier (2000 words/week - should be enough for most casual use), one time purchase for both local and cloud, and cloud with subscription ($6.99/mo) for when folks who need the latest cloud models. Not a fan of subscriptions too but hard to have ongoing support for the latest cloud models without it.
Learned some really cool things building this:
The interesting eng lesson for me has been that voice UX is so much more latency-sensitive than normal app UX - the major part of the work on this was on making it consistently low latency end to end.
On latency - the model is only one part of the delay. Shoute runs three backends for different modes and fallback (ElevenLabs streaming, Groq Whisper, and WhisperKit for on-device) and each has different latency profiles. For short recordings (~15s is my avg - Shoute can do really long but not the primary use case for hour long recordings), the annoying delays often come from everything around the model: audio finalization, connection warmup, WebSocket setup, token fetching, fallback paths, local model cold starts, and finally pasting into the active app. Getting all this right consistently took significant time and eng effort despite Claude helping with all of it - taste and architectural direction is still absolutely essential in 2026, especially with desktop and system apps.
Native development is still hard - things like WebSockets are fundamentally web technologies and their native libraries have a lot of hard edges and inconsistencies that only show up when you use something 100 times a day - took some engineering to get around this. Native does make the UX fast but it almost made me wish I had chosen Electron for something with this much network management, but speed and resource efficiency is worth going native for.
Okay, this already feels long - please try it, let me know how it feels, glad to hear feedback and feature requests. Thank you! Here is the link: https://getshoute.com/deepdive
Task duration is the time it would take for humans to complete the task. The speed of the models and how how long they might take to complete the task is not part of this metric.
It is the safest painkiller currently available. Ibuprofen can cause gut bleeding and renal issues if overdosed on. We all know about opiates.
Some facts - typical adult dose is 1g. Max suggested cap on the drug label is 3g per day (about 6 pills at usual 500mg dose). You need to take 10g (20 pills) to be at real risk of hepatotoxicity.
So 10 times the typical dose is when you have overdose effects. (basically 20 pills per day vs 2 pills per day).
Not your "wildly unsafe at slightly above usage levels" AT ALL (as someone posted on here)
This is not harmless - this might cause someone to take more dangerous painkillers when acetaminophen (tylenol) might have safely helped them. The autism stuff is plainly false and disproved.
> So that a delivery app, for example, can't strong-arm you into enabling location services by refusing to work otherwise
I can confirm this. I work for a top 10 consumer focused app (< #10 on the App store) and Apple did exactly this - ask us to enable users to use the app even without giving location permissions. There have been multiple push backs from Apple similar to this regarding user privacy and security and none from Google. And this isn't a vague generalization - Google has never asked us about this stuff, ever. It just isn't a priority. And so, as you might expect, I tend to trust and use the Apple platform for a lot of my personal stuff.
Apart from all the above: the fact that they are intentionally writing this (that they degrade frontier LLM dev, silently vs loudly for biology/cybersecurity) in the system card is interesting to say the least - especially just before IPO.
Notice that with this statement - that they're going to intentionally hobble the model for frontier LLM development - the general discussion has moved from, “Is the model actually that good?” to "they’re pulling the ladder up from behind them"
That's actually super smart - wonder if Mythos (or the next unreleased model) had a say in coming up with that strategy (if it's intentional). Also - having access to extremely capable models before anyone else - which they have by default - is a incredibly advantageous position to be in.