Cheap models orchestrate, frontier models get called only when needed
The top AI thread was 'A Jev that can ask Opus for help gives Opus accuracy at 4x lower cost' (50397429). Commenters immediately described their own versions. One uses a hybrid setup for security engineering work and rebuilt their harness's auto mode on the same idea. Another says Qwen3.8 27B can orchestrate and do shallow research, then escalate to Claude, and get answers close to Haiku, Sonnet or Opus. The obvious follow-up question was whether Sonnet or Haiku would work as the escalation target, which shows people are already tuning the routing layer.
Gemini Omni 1.2 Flash (50392118) landed the same day, and the first reaction was confusion over version numbers and a request for pelican-style tests. Model launches are now background noise. The routing logic around them is where people see real savings.
The pattern here: cost control has moved from picking one model to deciding when to escalate. Some commenters in older threads (48542100) also said local models like Qwen 3.6 27B feel roughly like Haiku 4.5 for coding, so the cheap tier is getting good enough to be the default.
So what?
If you sell on top of frontier APIs, your margin depends on how well you route, not which model you pick. Build an escalation layer now, and measure how often the cheap model can finish without help. If you compete on raw model access, a good router makes you easy to replace.