But getting a good result once isn't really the problem. How do you make this work for everyone in your org? People won't always have the right skill, and asking them to pick the right model isn't a great solution either.
That's what I looked at in this video.
I've been trying to figure out whether we can use cheaper models for knowledge work without losing quality.
So I compared DeepSeek 4.1 and Opus 5.5 on two weeks of our own calls, Slack and emails. Same task, with facts planted so we could check what they missed.
DeepSeek on its own wasn't good enough. Adding a skill helped a lot. Letting Opus lead and DeepSeek do the work through subagents got much closer, at about half the cost.
The most boring AI demo you'll see this week. On purpose.
Grokbot, Muse and Dot showed what a proactive AI agent does for one person. Workbot is the enterprise alternative: admins control what it reads, when it messages, and per-team rules.
We build for work, not media.
New in OpenWork ๐งต
More control for admins, and a smoother Workbot for everyone:
โข Jump to the latest reply
โข Workbot settings per team
โข Workbot cost analytics
โข Agent permissions
โข Website upload rules
โข Sign-in session controls
โข Personal API keys
A quick tour ๐
7/ Use OpenWork models in your own tools
Members can create a personal API key and call the OpenWork model gateway from scripts or tools like the OpenAI SDK.
Off until an admin turns it on. Existing model access still applies, and keys can be revoked anytime.
More control over file sharing. Set permissions for browser uploads, let members decide, and keep always-allow or always-block lists for specific websites.