ModelPar: find out if you're overpaying for AI

Workload	List	Paying	Could pay
Support agent claude-sonnet-5-5	$2.00	$1.36	$0.50 Fix prompt caching
Contract extraction gpt-6.1-sol	$2.00	$1.61	$0.81 Batch API
Call summaries gemini-3.8-flash	$0.75	$0.75	$0.25 Smaller model, to test

The problem

There's no public par for AI spend.

Your own bill is the only data point you see. That makes three problems hard to spot.

a.

List price is the only public number.

Every provider publishes a rate card. Discounts, commits and credits are negotiated one company at a time and kept private, so you can't tell whether your terms are good, typical or bad.

b.

Models get picked by habit.

Teams tend to use the model they started with, or whichever plan has limits left this week. Few test what each task needs, so simple work often runs on the most expensive model at the highest effort level.

c.

Bills and quality change silently.

A retry loop runs all weekend. A "latest" alias moves to a new version. Credits stop applying. A host serves a quantized model under the name you pinned. You find out from the invoice, or from your users.

How it works

Four steps. Start with a teardown.

Each step stands on its own, so you can stop after any of them.

01

AI bill teardown

A one-page report on what you pay, what you could pay, and where it goes.
You give

Usage and cost exports from each provider console. Seat counts and plans for coding agents. Your order form, if you have a commit.
You get
- Effective price per model vs list price, per million tokens
- Missed savings: prompt caching, batch pricing for async jobs
- Where spend concentrates, by model and workload
- Seat-vs-API math for coding agents
- Billing hygiene: hard limits, pinned model versions instead of "latest" aliases, credits that actually apply
- A routing card: task type, recommended model and effort level, cost per task, plan-limit headroom
Time

48 hours after we receive your exports
02

ModelPar Comps

An anonymized benchmark of what comparable startups actually pay.
You give

Your current terms: provider, discount, commit, term, credits and offer date. Opt-in.
You get
- Discounts by provider and commit band, with term, credits and offer date
- A band is never shown with fewer than 3 companies
Comps is new. Where a band doesn't have 3 companies yet, we say so instead of estimating.
Time

Shared with your teardown
03

Model and effort map

For each task type, the cheapest model and effort level that passes on your own work.
You give

For production: 30 to 50 representative tasks and what a good result looks like. For coding agents: 10 to 20 recently merged PRs.
You get
- A map from task type to model and effort level, scored on quality, cost and latency
- The same task set run nightly as a canary, to catch silent quality drift from quantization or version changes
- A refresh whenever a new model ships
Your tasks, your choice

Run production tasks in your own environment, and only scores, token counts and timings leave it. Or share them with us, so we can read the failures, tune the grading with you and explain why a model fell short, not just that it did.

How the PR replay works

Each PR is replayed per model and effort level on your machine. Only scores, token counts and timings leave it.

Time

2 to 3 weeks for the first map, then nightly
04

Negotiate at renewal

A competitive process across labs and inference providers, for teams spending $250K+ a year.
You give

Your current contract, usage history and renewal or commit date.
You get
- Competing offers from labs and inference providers, checked against Comps
- Your model map, showing which workloads can credibly move
- You make the final call on every term
Time

Start 60 to 90 days before renewal

Sample report

One page you can forward to your CFO.

Every teardown uses the same format: dated, versioned and specific to your bill. This one was prepared for a fictional 14-engineer company.

AI bill teardown

Confidential

Example Co.

14 engineers · B2B support software

Version: v1, prepared for Example Co.
Date: Nov 3, 2026
Billing period: Oct 1 to Oct 31, 2026
Sources: Anthropic Console cost export · OpenAI usage export (cost data) · Google Cloud Billing report · AWS Cost Explorer (Bedrock) · Together AI cost analytics · Claude Code analytics · Cursor team analytics
Prepared by: ModelPar, [email protected]

Summary

October spend	$26,234 ($315K a year at this rate)
Savings with no model change	$6,854 a month (26%)
More if one model swap passes testing	$2,067 a month
Coding-agent seats	$2,158 a month by moving 4 engineers from API billing to Team seats
Billing hygiene	5 fail · 1 warning · 3 pass

Two workloads are 51% of spend: the support agent (26%) and nightly contract extraction (25%). Both have a fix that doesn't change the model.

1. Monthly spend by provider and model

October spend by model

claude-sonnet-5-5 $9,096
gpt-6.1-sol $6,500
claude-opus-5-5 $4,718
gemini-3.8-flash $2,813
claude-sonnet-5-5 (Bedrock) $1,538
claude-sonnet-5-5 (in Cursor) $623
DeepSeek-V4.1-Flash $616
Cursor seats $270
text-embedding-3-small $62

Workload	Provider	Model	October	Share
Support agent	Anthropic API	claude-sonnet-5-5	$6,920	26.4%
Contract extraction (nightly)	OpenAI	gpt-6.1-sol	$6,500	24.8%
Call summaries	Google Gemini API	gemini-3.8-flash	$2,813	10.7%
Claude Code, API key (5 engineers)	Anthropic API	claude-opus-5-5	$2,610	9.9%
Ticket triage	Anthropic API	claude-sonnet-5-5	$2,176	8.3%
Support agent, EU customers	AWS Bedrock	claude-sonnet-5-5	$1,538	5.9%
Evals and staging	Anthropic API	claude-opus-5-5	$1,152	4.4%
Escalation drafts	Anthropic API	claude-opus-5-5	$956	3.6%
Third-party model usage, incl. Cursor Token Rate	Cursor	claude-sonnet-5-5	$623	2.4%
Search query rewriting	Together AI	DeepSeek-V4.1-Flash	$616	2.3%
Coding-agent seats	Cursor	Teams, 9 Premium seats	$270	1.0%
Embeddings	OpenAI	text-embedding-3-small	$62	0.2%
Total			$26,234	100%

By provider: Anthropic API $13,814 (52.7%) · OpenAI $6,562 (25.0%) · Google Gemini API $2,813 (10.7%) · AWS Bedrock $1,538 (5.9%) · Cursor $893 (3.4%) · Together AI $616 (2.3%).

Read the full sample report

Your data

We need billing data, not access.

Exports, not keys.

The teardown starts from usage and cost exports you download yourself. No API keys, no console access.

No prompts, code or customer data, unless you choose.

Usage exports hold token counts and dollars, not content. The teardown and Comps never see prompts, completions, code or customer records. For the model map, sharing production tasks is your call.

The PR replay runs on your machine.

Only scores, token counts and timings leave it.

Deleted on request, or after 90 days.

Your files are deleted when you ask, and automatically 90 days after we receive them.

Benchmark contributions are opt-in.

Your terms join Comps only if you choose. They're stored without your company name, and no band is shown with fewer than 3 companies.

Mutual NDA available.

We're happy to sign one before you send anything.

Some exports include API key names, project names or user emails. You can delete those columns before sending. We don't need them.

Who's behind it

Joe Waltman

Founder

Joe co-founded VetPronto, a mobile veterinary care company, in Y Combinator's W15 batch. Before that he co-founded RestEngine, which was acquired by Twitter. Most recently he was a Group Product Manager at Coinbase, from 2018 to 2022.

He started his career as a network engineer at Qualcomm, and holds an MS in Computer Science from USC and an MBA from INSEAD.

We're early and working with our first 10 teams.

LinkedIn [email protected]

FAQ

Questions

How is this different from AI spend dashboards?: Dashboards show where your money went. We tell you what you should be paying and which model each task needs. We work alongside them: if you already use one, its export is a fine starting point for the teardown.
Do you need API keys?: No. The teardown and Comps work from usage and cost exports you download from each provider's console. You never hand over keys or console access to get started.
Do you resell credits?: No. We don't resell credits or tokens.
How is my data protected in the benchmark?: Contributing is opt-in. We remove your company name before your terms are stored, group them by provider and commit band, and never show a band with fewer than 3 companies. You can withdraw your contribution at any time.
Do you see my code?: No. The teardown uses billing exports, which contain no code. For coding agents, recently merged PRs are replayed on your machine, per model and effort level, and only scores, token counts and timings leave it.
Do I have to share prompts for the model map?: No. You can run the task set in your own environment, and only scores, token counts and timings leave it. Sharing tasks with us is optional. It lets us read failures, tune the grading with you and explain why a model fell short, not just that it did.
What if there are no savings?: Then the teardown says so, and you have a dated record that your setup is sound.
Which providers do you cover?: OpenAI, Anthropic, Google (Gemini API and Vertex AI), AWS Bedrock, Azure, open-model hosts such as Together AI and Fireworks, and coding-agent seats including Claude Code, Codex and Cursor.
What do I need to send?: The last 3 full months of usage and cost exports from each provider, your seat counts and plans for coding agents, and your order form if you have a commit. After you request a teardown, we send links to each provider's own export instructions.

Find out if you're overpaying for AI.