The decision gets made once
A model is chosen the day a feature ships and stays chosen. Prices move every month. Nothing in your process is watching.
“We chose that model 14 months ago. It's probably wrong now. Nobody's checked.”
For finance & engineering leadership
Cut 15 to 40% of your LLM bill without cutting quality.
You are overpaying for AI, and we can prove exactly where. Add one line of code and get a full report card on the price and performance of your AI application.
- Every route priced per call, then replayed against cheaper models on your own provider accounts. The savings come back quantified in dollars, not adjectives.
- Nothing is recommended until quality provably holds. We are never in your request path.
- The change arrives as a pull request with the evidence attached, so the saving lands in weeks, not quarters.
Your LLM bill today
Recovered on a $50k/month bill
Save $90k to $240k / year
Same quality, proven on your own traffic before anything ships.
The problem
Nobody owns the question “are we running the right model?”
Picking the right model is a laborious, time-consuming process that requires auditing logs, building evals and running tests.
Known savings sit in a backlog
Most engineers can point straight at the waste. It's two weeks of work, which is exactly why it loses to revenue work every quarter.
“Free money. It still hasn't made the list.”
The downside risk stops everyone
Switching without evidence bets product quality on a hunch. So the expensive model wins by default, and keeps winning.
“Pull one lever, improve five characteristics, degrade fifteen others.”
Quotes from customers spending $10k to $200k+/mo on LLMs
The shape of the deal
A real cut to the bill that barely touches your engineers.
15 to 40%
Modelled spend reduction
What we model once the routes worth optimizing have been optimized.
One afternoon
To be fully instrumented
Wrap the client your team already uses. By the end of the day you can see what every AI call cost and which feature spent it.
Every new model
Tested against your traffic
Checking whether a new release beats what you already run is a project nobody schedules. We run it every time, so you keep the cheapest model that holds quality.
the 15 to 40% figure is our model, not an audited customer result
Where the money is
Model swaps get the attention. Most of the money is in the work nobody schedules.
Each of these is its own project, which is why each one keeps getting deferred.
| Lever | Typical reduction | Why it hasn't happened yet |
|---|---|---|
| Model substitution | 40 to 70% on the route | Needs an eval set nobody maintains, and a quality case nobody can win without data. |
| Batch API migration | ~30% where eligible | Two weeks of unglamorous work that loses to the roadmap. |
| Prompt distillation | 50 to 80% on the route | Takes a pipeline that teaches a cheap model what the expensive one knows. |
| Triage routing | 30 to 60% blended | Safe only once you can measure what the cheap model handles at parity. |
| Parameter tuning | 10 to 30% | Reasoning budgets, retries, and fallbacks are invisible without per-feature cost. |
| Cache strategy | 10 to 25% | Shows up on no dashboard you own. |
| Feature cost drivers | varies, often large | When a provider changes how a feature bills, you find out on the invoice. |
Per-route ranges, applied only to the slice of your traffic that qualifies. Not a promise about your total bill.
The spread that makes this worth testing
Below the frontier tier sits a set of models costing a fraction as much per token, and on a lot of routes they hold up fine.
| Model | Input / 1B | Output / 1B | Cost index |
|---|---|---|---|
| Frontier tier · where most production traffic sits today | |||
| Claude Fable 5 | $10,000 | $50,000 | 200 |
| GPT-5.6 Sol | $5,000 | $30,000 | 114 |
| Claude Opus 5 | $5,000 | $25,000 | 100 |
| Where a lot of that work runs just as well | |||
| Kimi K3 | $3,000 | $15,000 | 60 |
| Gemini 3.5 Flash | $1,500 | $9,000 | 34 |
| GLM-5.2 | $760 | $2,420 | 11 |
| DeepSeek V4 Pro | $435 | $870 | 5 |
| DeepSeek V4 Flash | $90 | $180 | 1 |
Cheaper per token is not the same as cheaper per task
A model that reasons longer or retries more can cost more per finished answer than the one it replaced. This table is the reason to look, not the answer.
The proof standard
Cheaper is only a saving if quality holds. So we prove it first, in writing.
The objection your CTO will raise, answered
Any vendor can tell you to run a cheaper model. Teams don't, because a quiet drop in quality nobody catches for a month costs more than the savings.
So we replay your real production inputs against the challenger and score every quality characteristic with a confidence interval. A recommendation appears only once that interval clears your parity threshold, and it still has to survive a live canary.
−2% margin0+2%
accuracy
+0.3
faithfulness
+0.1
format validity
+0.0
tool-call correctness
+0.4
tone
−0.6
illustrative scorecard · judge reliability on this route: 0.87 agreement with your labels · bias-probed · sampling stopped at CI convergence
How it works
Four steps. Your team's involvement is step one and a code review.
-
STEP 01
Measure
One line of code attributes every AI call to the feature that made it, in dollars. Usage metadata only.
-
STEP 02
Auto-eval
Eval sets are generated from your own traffic, never hand-written, then replayed against cheaper models inside your allow-lists and residency rules.
-
STEP 03
Prove
A scorecard per feature: quality parity with confidence intervals, plus cost and latency deltas.
-
STEP 04
Ship
A pull request with the evidence attached, rolled out behind a canary, reversible with one flag.
The risk question
What you are actually exposing by trying this.
It cannot take down your product
A library, not a gateway. Capture is asynchronous, and none of your traffic is proxied through us.
Your customer data stays with you
The SDK sends usage metadata only. The open-source server strips content at ingest, so prompts have no path into the database.
Regulated environments are supported
Testing runs on your own accounts, inside your allow-lists and residency rules. We also deploy inside your VPC, sign a BAA and handle PHI under HIPAA, and provide SOC 2 evidence for your security review.
Quality is proven before anything ships
Non-inferiority testing, then a live canary, then promotion. Every step reverses with a flag.
Leaving is a one-line change
Apache-2.0 and yours to self-host. Everything exports, and switching off means editing one environment variable.
You can start without procurement
The tracking layer is free and runs on your own infrastructure. Size the opportunity before you talk to us.
FAQ
What leadership asks us first.
Our engineers say they've already optimized this.
Once, on the routes they had budget for. Whether that answer survived six months of price changes is a different question. Run the free layer for a week and you'll know.
What if a cheaper model quietly degrades our product?
Nothing is recommended until the challenger clears non-inferiority testing against your own threshold, judged by a model qualified against your labels. Then it has to survive a live canary, and one flag reverses it.
How much engineering time does this cost us?
An afternoon to wrap your existing client. After that your team reviews pull requests that arrive with their own evidence.
How fast is payback?
Cost per feature shows up within days of instrumenting. Change recommendations follow once there's enough traffic to reach confidence: weeks on high-volume routes, longer on quiet ones.
What does it cost?
The measurement layer is free and open source forever. The hosted engine is a platform fee tiered on traced volume, free for design partners while we calibrate. If the modelled savings don't clear the fee at your volume, we'd rather tell you.
We already have an observability or gateway vendor.
Metergraph sits alongside both. Observability records what your spend was; a gateway routes by rules a human wrote. Neither runs the experiment, and neither opens the pull request.
What about compliance, residency, and sensitive data?
Defaults are metadata-only capture, your own API keys, per-route opt-outs, zero-retention modes, and residency-aware sampling.
For stricter obligations we deploy inside your own VPC, sign a BAA and handle PHI under HIPAA, and provide SOC 2 evidence for your security review.
Where does the 15 to 40% number come from?
It's our model, not an audited customer result: the share of spend that typically sits on optimizable routes, times the per-route ranges above. Your number depends on your traffic mix.
Next step
Find out what your number actually is.
Tell us roughly what you spend and we'll scope an assessment against your real traffic.
$60k / month
Estimated annual savings
$108k to $288k
Or start measuring yourself · open source on GitHub ↗ · read the docs