Accel AI Innovate · Amsterdam · 23 September 2026
Route
Find out what your agents cost.
Then make them cost less.
One command reads your whole agent estate and shows what can move to open weights on Nebius Token Factory — with the measurements behind each one.
ax-route.vercel.app · github.com/lpsmurf/ax-route
Who this is for
You always pick the best model. Nobody ever told you which jobs don’t need it.
0.13%
of my requests used the cheap model. I had even written down that they should.
If you build with AI all day and you are not an infrastructure engineer, you are almost certainly doing this.
Opus for everything. Including “does this file parse”.
Nobody is going to hand-tune this. It has to be measured for you.
measured · 3,865 deduped requests, 50 days · results/replay.json
What you are overspending
$275.35
of your $670.73 is overspend — same work, cheaper model
4,005 requests · 56 projects · 8 agent roles
| project | paid | overspend |
|---|---|---|
| C-01 | $259.81 | $111.89 |
| SA-02 | $127.51 | $48.98 |
| VDR-03 | $69.6 | $30.15 |
| W-04 | $23.37 | $9.21 |
labels, not client names · route audit --reveal
measured on this machine · results/estate.json
Why this is only possible now
Two years ago the cheap tier could not do the job at all.
2024
Small open models failed mechanical work outright. Routing had no upside, so nobody built the governor.
2026
Open weights cleared the bar. Nebius Token Factory puts 24 of them behind one endpoint at a published per-token price you can compute against.
The gap is now hundreds of times — and the quality difference is noise.
prices read live from the Token Factory API · GET /v1/models?verbose=true
What you actually do with it
One command. Then one line to change.
Step 1 — see it
npx ax-route audit
No key. No account. Reads the logs your tools already keep.
Step 2 — change it
~/.claude/agents/memory-retrieval.md
- model: opus
+ model: haiku
Or point Cursor at Route and it switches by itself.
It tells you which files to change, and why.
And it refuses to tell you about the other five jobs, because it has not tested them.
Every routed call leaves a receipt: model, tokens, cost, region.
Every figure here traces back to a logged call you can open.
ax-route.vercel.app · the dashboard is the demo
The routing choice, measured
I picked a model by hand. The measurement overruled me.
All 19 models on Token Factory, 21 real tasks, one blind judge. The most expensive model tested scored the worst.
Accuracy 66.7–95.2%. Price 172×. Paying more buys you nothing here — which is exactly why you measure.
Against the models you actually pay for — measured
An open model matched Opus exactly. It cost 159 times less.
| model | judge pass | cost, 21 tasks | p50 |
|---|---|---|---|
| google/gemma-3-27b-it (open, Nebius) | 95.2% | $0.000319 | 378 ms |
| claude-haiku-4-5-20251001 (Anthropic) | 90.5% | $0.003697 | 814 ms |
| claude-sonnet-5 (Anthropic) | 81% | $0.013938 | 1693 ms |
| claude-opus-5 (Anthropic) | 95.2% | $0.050645 | 2564 ms |
Real calls to the Anthropic API on the same 21 real tasks, graded by the same judge, blind to which model answered. The cheap open model at the bottom of that table scored 95% on toy questions and 81% on these. Earlier versions of this deck had to label this comparison computed. It is not any more.
measured · results/frontier.json · whole run cost $0.0683
What it is worth
Replayed against 50 days of my real work.
3,985 real requests, re-priced from raw token counts: $664.87.
| share moved | bill becomes | saved | saved % |
|---|---|---|---|
| 25% | $528.43 | $136.45 | 20.5% |
| 50% | $391.98 | $272.89 | 41% |
| 75% | $255.53 | $409.34 | 61.6% |
A band, not a point — the mechanical share is an assumption. I first published $164 here. My own code billed cached tokens at zero, and cache reads outnumber fresh input fifty to one. Fixing it made the problem four times bigger.
measured baseline, scenario projection · results/replay.json
Business model
We get paid a share of what we actually save you.
Free · the hook
The audit
One command, no key, no account. It hands you a number about your own machine. That number is the marketing.
Paid · the core
20% of verified savings
We already measure the before and after per call. You only pay out of money you were going to burn anyway.
Teams
Control plane
Shared policies, spend caps, per-project ledgers, approvals. Seat-based, for the people who have to sign off on the bill.
How it spreads
Every run ends with a figure people want to post. npx ax-route audit costs nothing to try and needs no sign-up, so it travels through the Discords and timelines these users already live in. They land as individuals; their teams follow.
Why it is worth paying for
They are overpaying and cannot see where. Nothing else reads their own machine and proves which work is safe to move. And Nebius serves from Finland, so the same product answers the European question: this code never leaves the EU.
Reselling inference at a margin would pay us more when you spend more. We would rather be paid when the bill goes down.
Every number names the file it came from.
Built today
- OpenAI-compatible router, zero dependencies
- 19-model sweep on Token Factory
- Blind-judged benchmark, n=20
- Replay over 3,865 real requests
Not built, and said so
- No measured frontier arm — no key
- Escalation gate is designed, not shipped
- Claude Code needs an Anthropic adapter
- Fixtures authored, not sampled
ax-route.vercel.app
github.com/lpsmurf/ax-route · MIT