I haven't used Opus 5 yet. I'll take it for a spin this week, same as everyone else who's been reading the arguments instead of running the benchmark themselves. But I already know what I'm going to find, because I already know how I actually work: for the price, I do better managing Sonnet than reaching for the frontier model by default. I wait until I see the cracks in the model I'm using, and only then do I have it prep the problem for something better.
That's not a workaround. That's the job.
The Debate Is Happening One Layer Too High
Ars Technica's framing of Opus 5 — a token-efficiency update, not a capability leap — is correct, and the discourse split into exactly the two camps you'd expect: people disappointed it wasn't a bigger jump, people celebrating the cost savings. Both camps are debating the model. Neither is debating the thing that decides whether the model gets called.
The question was never which model. It's which layer you're arguing at.
Here's the layer most of that discourse skips entirely: the one that decides, before a single token reaches Opus 5 or Sonnet or anything else, whether this request even needs the expensive model — and if it does, what it should be told to do once it gets there.
Routing Is Not One Decision. It's Two.
Most "smart routing" talk stops at model selection: cheap model for easy tasks, expensive model for hard ones. That's real, but it's half the picture. My own pipeline — Fire, the API layer behind everything I build on — routes on two axes, not one: which model handles a given call, and what system prompt it runs under. Two requests can hit the identical model and do completely different jobs because the prompt riding underneath them is different.
That second axis is where the actual leverage lives. Right now I can call a flow that sends one prompt to two "hot" agents in parallel — one of them running under a system prompt whose entire job is to disagree — and route both results, plus the original prompt, into a third "cool" agent for synthesis. Three model calls, one answer, and the disagreement is structural, not something I had to think to ask for. None of that required a smarter model. It required a system prompt that knows to argue.
The obvious next move — and it's a small one from where the pipeline already sits — is putting a fast, cheap model in front of a flow like that whose only job is deciding which models the flow underneath it should actually use. Not routing the request. Routing the router.
The Stack Underneath the Argument
Once you're routing on two axes and nesting flows inside flows, you're not really talking about "which model" anymore. You're talking about a stack, and the model only shows up in the bottom third of it.
The Stack Most Efficiency Debates Skip
4. UI — where the request enters.
3. Threading & decomposition — the prompt gets taken apart, gets a thread ID, personality gets injected, and results downstream get synthesized back into one answer.
2. Routing & queuing — which model, which system prompt, timeout handling.
1. Model and tool calls — the actual inference, the actual database hit.
Opus 5 lives entirely in layer 1. A token-efficient model at layer 1 makes every decision made at layer 2 cheaper to be wrong about — the cost of escalating too early drops, so the threshold for "wait until you see cracks" can sit lower without the bill climbing. That's the actual value of this release. It doesn't make the model smarter. It makes the layer above it braver.
Fix the Layer You're Actually Arguing About
I said something close to this before, about a different stack: the discourse fixates on the middle floor — better prompts, more context, chain-of-thought — while the floor underneath leaks. Same failure mode here, one level down. The Opus 5 debate is entirely a layer-1 argument dressed up as a strategy question. The strategy question is layers 2 and 3: what decides when layer 1 gets called, and under what instructions.
Take Opus 5 for a spin this week. Just don't mistake the spin for the decision that actually saves you money.
— J.P. Howlett
Related:
- The Meter Runs — the "physics, not a pricing choice" argument this piece routes around: efficiency at layer 1 is what buys you room to be braver at layer 2.
- Three Tiers — the wrong-floor argument this piece borrows its shape from, one level down the stack.
Sources