Opus 5.5 with Grokbot

So... this is how I run Opus 5.5 with Grokbot across my bot fleet.

I've got a lot of bots doing a lot of different jobs. Some need deep thinking. Some need a picture, a voice, or a talking-head video. Sending all of it to one big model is like asking your best engineer to answer the phones too.

So there's a gatekeeper out front. It's called Jev. It's TypeSafe's "System One" decision model.

Jev doesn't write anything. You hand it the job and a short list of buckets: hard reasoning, image, lip-sync video, voice, document. In about a tenth of a second, it hands back a pick, and how sure it is. Just probabilities. And it's very cheap.

My Grokbot fleet already leans on Jev gates. Should this routine wake up? Who takes this job? Is it ready to release? Should we stop retrying?

Once Jev picks the bucket, a plain routing table takes over.

Hard reasoning goes to Anthropic's Opus 5.5. That's the primary model for the tough stuff. If Opus can't take it, Grok 4.7 is the fallback.

Images go to Grok Imagine Image 2.0.

Lip-sync and talking-head video go to Grok Imagine Video 1.5, with Hedra as the backup.

And voice goes to ElevenLabs, on my own voice. You're hearing it now.

There's one more lane. Some client data has to stay in the US. Think of a mining client. That work goes to xAI's US-only endpoint. It costs about ten percent more, and for that data, I'll pay it.

Now, the safety valve. When Jev isn't sure, Jev doesn't decide. If its confidence is low, a person makes the call, or the main model does.

So the quick, cheap calls do the sorting, and Opus gets saved for the hard thinking.

I've watched a few of these waves come through. The tools that stick make good people faster... and they know when to hand it back.

That's Opus 5.5 with Grokbot. Jev at the front door... and a person on the close calls.