Metergraph Savings How it works Tools Docs Pricing Blog Start saving now

An optimization agent for your AI

Cut LLM costs without cutting quality.

Metergraph reads the traces your AI agents already produce, finds where you overpay for intelligence or lose quality, tests fixes on your own traffic, and helps you ship them.

Start saving now → Get your savings review

80% lower AI costs with better answers for one design partner. Half the optimization work and 15% lower costs for another.

Free up to 100K calls a month. Already have an account? Sign in

Savings found · example
/mo

Shorter prompt for Rank competitive signals

Quality
+1.0%
Latency
1.61× faster
Tested on
20 real requests

Lower AI spend

Reduce AI spend without risking quality.

Metergraph tests each change on your production traffic first, so you see what it saves and how it affects answer quality before you approve it.

80% lower AI costs for a design partner with unoptimized traffic, with better answers
50% less engineering time on optimization for a partner who already optimized regularly, and 15% lower costs on top

Paying more for intelligence does not get better answers. A newer model from another provider cost far less and hallucinated less on their own traffic. Read the story ↓

Ready to ship

Every recommendation arrives tested on your real requests and graded against what you ship today.

Savings you keep

You pay a flat rate per call. We never take a percentage of your spend or of your savings.

Numbers you can stand behind

Each saving links to the requests, answers and grades behind it.

Import your existing traces and see the opportunities we find. Most teams are surprised.

Start saving now → Get your savings review

Better product, same team

Do you know what your product spends tokens on?

Save money, improve quality, and keep your team on product work. Here is how Cascade Marmot, an example company spending about $12,700 a month on LLM calls, finds out.

Performance groups spend by workflow. One workflow, Rank competitive signals, accounts for 85% of the $12,720 spent in the last 30 days.

Metergraph Performance page: spend, calls, error rate and p95 latency for 30 days, with daily spend grouped by workflow

Read the full walkthrough →

Works with your coding agent

Metergraph's optimization agent learns how your users use your system, then finds changes and tests them on that traffic before anyone opens a pull request, so problems are caught upstream. Your coding agent, such as Claude Code, Codex or Cursor, makes the change, and your team reviews the pull request.

A model trained on your traffic

When no off-the-shelf model fits a workload, Metergraph can train one on your own traffic and compare it with the models you use today before you switch.

Boost your team →

Tools for builders

The tools you'd build yourself, ready to use.

Benchmarks built from your own usage, cost tracking by workflow, model experiments on your traffic, and skills for your coding agent.

Free, no account

Pricing API

Know what every LLM call costs. Send a provider response and get the exact cost, with prompt caching, batch discounts and reasoning tokens counted in.

curl https://app.metergraph.dev/v1/pricing/calculate \
  -H 'Content-Type: application/json' \
  -d '{"format": "anthropic.messages",
       "trace": {"model": "claude-sonnet-5",
         "usage": {"input_tokens": 1000, "cache_read_input_tokens": 800,
                   "output_tokens": 200}}}'
Pricing API docs →

Open source

Skills for your coding agent

Install Metergraph's skills and your coding agent knows how to find where you overspend, compare models on your traffic and propose the change. Works in Claude Code, Codex and Cursor.

claude plugin marketplace add metergraph/skills
claude plugin install metergraph@metergraph
metergraph/skills → 
$ pip install metergraph

import metergraph
from openai import OpenAI

client = metergraph.wrap(OpenAI())

Free up to 100K calls a month

Every feature is included, and it covers most side projects.

Open-source dashboard

Run the Apache-2.0 spend dashboard on your own machine. GitHub →

More free tools

Estimate model costs with the cost calculator, or compare four models on your own prompt in the blind test.

Does being rude to AI cost more? Try the tone experiment →

Start building free → Read the docs →

Pricing

Free to start, with flat pricing as you grow.

Free

$0

100K calls a month with every feature.

Start saving now

Growth

$399/month

250K calls included, then $99 per additional million. Same features as Free.

Start saving now

Scale

Custom

Deploy in your VPC or on-prem, with SSO. Our team runs the analysis and ships the pull requests with you.

Get your savings review

Compare plans →

Start with the traces you already have.

Import them and see the opportunities we find. Most teams are surprised.

Start saving now → Get your savings review

Not ready to sign up? Leave your email and we'll be in touch.

Design partner stories

Paying more for intelligence does not get better answers.

An AI-native product team's spend was concentrated in a few functions. Metergraph recommended switching them to a newer-generation model from another provider.

The new model cost significantly less and hallucinated less, and the evidence report showed both on the team's own traffic.

80%lower AI costs, with better answers

Already optimizing? Half the work.

Another design partner tuned its AI workflows regularly, by hand. With Metergraph finding and testing the changes, the team spent half as much engineering time on optimization and still lowered costs by another 15%.

50%less engineering time on optimization
15%lower costs on top of their own work

When it's time for Metergraph

A model you rely on is being retired

Replay your traffic on the candidates and switch on evidence.

A new model or a price cut shipped

Find out whether it does your work as well, at a lower cost, before you switch.

Retries and failures are adding up

See which workflows fail, what the retries cost, and which setup holds up.

Prompts keep growing

Find the workflows that carry more context than they need, and what trimming saves.