Everyone selling an LLM router will tell you it saves 60% on your AI bill. Nobody shows you the numbers on a real workload. So I pulled my own.
Over the last three months I've run coding agents hard, Claude Code mostly, day in and day out. Every request it makes is logged locally with its token counts, so I parsed the lot: 1,231 sessions, 1.25 GB of logs, 139,835 billed requests, 1 June to 25 August. Here's what a real agent workload costs, and what routing would do to it.
The raw number
39.5 billion tokens. At Anthropic's API rates that's about $29,962, call it $10,000 a month.
Two things jump out.
First, 91% of that cost is Opus. Almost every request went to the biggest, most expensive model, whether it needed it or not. A one-line rename, a test refactor, a "summarise this diff", all of it to the frontier model at frontier prices. That's not a criticism of me. It's the default, and the default is expensive.
Second, and this surprised me, the bill is dominated by context, not output. The single biggest line is cache-read tokens, the whole conversation re-billed on every turn. By the hundredth turn of an agent session you're paying for the entire codebase again to add one function. Output is the small part.
What routing does to it
The pitch for routing is simple. Send the work a cheaper model can handle to a cheaper model, and escalate to the frontier only when you need it. The question is how much of the work is genuinely cheap-model-able.
My companion benchmark, OmnisBench, measures that on fresh coding problems the models can't have memorised. There, ideal routing matches the frontier model's quality at roughly 60% lower cost, because most coding requests don't need the biggest model.
Apply that to my $30k. Route the 60% of requests a cheaper model handles down to a small model, keep the rest on Opus, and the equivalent bill drops to somewhere between $12,500 and $15,600. A 48 to 58% cut. About $5,000 to $6,000 a month, on a $10,000 workload. That lands right on the benchmark's number.
The honest bits
Two things I won't paper over.
I didn't pay $30k. I ran this through a Claude Max subscription, which already saves something like 98% versus the API. So this isn't a bill I paid, it's a realistic workload standing in for a team that does pay API rates. If your team runs agents on your own keys, this is roughly your shape.
And the reprice is modelled, not proven. The 60% comes from the benchmark's aggregate, not from re-running every one of my requests through a cheaper model to confirm it passed. Proving that costs real API money, and it's the honest next step, not a number I'm going to pretend I already have.
Why I trust the direction anyway
Because there's now an independent check. RouterArena, an ICLR 2026 benchmark, ranks routers on accuracy against cost, and it found the loud commercial ones frequently over-pick expensive models. It's the neutral scoreboard the field didn't have. OmnisRouter is built for exactly what it measures: pick the cheapest model that's still correct, escalate when unsure, and show the receipt.
Do it yourself
The whole point is that you don't have to take my word for it. Your Claude Code logs are on your disk, with the token counts, in ~/.claude/projects. Parse them, price them at API rates, and see your own split. Mine was 91% Opus. I'd bet yours is close.
The three pieces, all open:
- OmnisBench measures which model each kind of work needs, on fresh problems, re-gradable offline.
- OmnisRouter routes each request to the cheapest capable model, with a receipt on every response.
- OmnisVigil rolls it up per team, so a lead sees the runaway agent before the invoice does.
Tell me what your split looks like.
Links: github.com/Fortitude-Group/OmnisRouter · omnisbench.fortitude-omnis.group · omnisvigil.fortitude-omnis.group
Top comments (0)