3000 tokens/s : Welcome to Alter Fast
17x faster than GPT-5
Samuel ROY & Olivier Legris
October 10, 2025
Hey everyone,
We are launching a new AI router: Alter Fast
Based on your needs, we will intelligently route your request to the right model from a selection of the fastest available, our fastest one being 17x faster than GPT-5 high.
It’s more than a performance toggle, it’s a re‑imagining of what AI UX can be.
We are so confident people will love it, it is now the setting per default for all new Alter users.
What's New?
We handpicked the fastest models available on the market, without sacrificing quality, with the fastest one peaking at 3000 tokens/s.
Feel free to try that one individually too, it is Cerebras/gpt-oss-120b
Be aware, speed is addictive.
How to enable it
When typing / to display your model list, you can now see Alter Fast; just select it and you can drive on the high‑speed lane of information.
Should you I use it Fast all the time?
We’ve been testing fast for a week now, and for the vast majority of the tasks, it works like a dream.
If you are a heavy user of tools, you may still want to keep /best, Claude 4.5, or GPT‑5.
Also, don’t forget you can specify which model you want to use at an Alter Action level in the Advanced tab.
If I Had More Time,
I Would Have Written a Shorter Letter
Your feedback are important!
As we’ve been playing with this for less than a week, I’m sure we have not faced the diversity of tasks and contexts. So let us know if you find a scenario where /fast fell short.
And of course, feel free to share some love too, if you like it.
Cheers
Full Changelog
New Features & Enhancements
Models: Introduced a fast Alter model pipeline with smart auto‑selection and 3000 tokens/second
Notch Cursor: Added proper text field keyboard navigation in the notch
Dictation: Now replacements respect word boundaries
Bug Fixes & Stability
Notch & Windows: Kept Alter windows in front when closing the notch
Infrastructure: Optimized load balancers configuration and grace periods during streaming on sigterm signals
Conversations: Multiple crash fixes, fixed the invisible draggable area while minimized and better support of streaming finish reasons
Models & Endpoints: Fixed custom endpoint where an empty URL produced an empty model list
Audio & Recording: Mitigated a double‑free audio crash and the recording clock
Core & Startup: Improved error reporting when Core Data fails and prevented crashes on MCP Client start errors