Claude Sonnet 5.5

I learned through my testing that sonnet 5.5 sometimes spawns opus 5.5 subagents. Might be one of the main reasons why sonnet 5.5 feels a lot more expensive for people.

And, this is not about the API price, but the percentage weekly usage of the subscription.

I actually added a mini router in front of it this week, and the harder slices get dispatched to sonnet 5.5 instead, pretty effective so far, it improved first-pass implementation quality.

So I don't know what's your process but I would suggest models wise to just use Opus 5.5 high or medium as the orchestrator to supervise a Sonnet 5.5 high implementer and use 6.1 sol xhigh (or medium if you notice xhigh is not sustainable for you) as a reviewer.

It really depends on the task, but it's often either sonnet 5.5 high or opus 5.5 med.

I've never once hit a safeguard refusal. Dunno what you guys are doing

If Astra is that great, how come its code introduces so many bugs? Give Astra a complicated task on Ultra, and then after give the prompt "review" - you will get about 5 problems.

Tokens | cost to serve/run the same agentic task | performance across multiple benchmarks