Catch AI API cost regressions before they ship

One verified dataset of official LLM API prices, workload cost computed from it, and a budget CI enforces.

This is what a one-line model change looks like to SpendLint. The pull request swaps Claude Sonnet 5 for Claude Opus 5 on a workload of 1,000 requests a day. The contract below allows $500 a month and a 20% rise, so the check fails and names the rules that broke.

Change	Workload	Model	Tokens per request	Per request	Per month
−	Baseline	Claude Sonnet 5	3,000 input, 600 output	$0.01	$360.00
+	Proposed	Claude Opus 5	3,000 input, 600 output	$0.03	$900.00
=	Delta per month (+150.00%)FAIL				+$540.00

1,000 requests per day, 30 day monthPrices effective 2026-10-22Verified 2026-09-05Source Anthropic

Rules violated: max_monthly_usd, max_cost_per_request_usd, max_regression_pct. 5 cheaper models satisfy the contract. Check a contract with your own numbers.

Flagship model prices today

USD per 1 million tokens, standard tier, read from each provider's official price page.
Model	Input per 1M	Output per 1M	Verified	Full table
Claude Sonnet 5	$2.00	$10.00	2026-09-05	All Anthropic prices
GPT-5.4	$2.50	$15.00	2026-09-05	All OpenAI prices
Gemini 3.1 Pro Preview	$2.00	$12.00	2026-09-05	All Google (Gemini API) prices
GPT-5	$1.25	$10.00	2026-09-05	All Azure OpenAI (Foundry Models) prices

Cheapest current models for a 3,000 in, 600 out workload

Model	Provider	Per request	Per month at 1,000 req/day
GPT-5 nano	Azure OpenAI (Foundry Models)	$0.000390	$11.70
GPT-5 nano	OpenAI	$0.000390	$11.70
GPT-4.1 nano	Azure OpenAI (Foundry Models)	$0.000540	$16.20
Gemini 2.5 Flash-Lite	Google (Gemini API)	$0.000540	$16.20
GPT-4.1 nano	OpenAI	$0.000540	$16.20
GPT-4o mini	Azure OpenAI (Foundry Models)	$0.000810	$24.30
GPT-4o mini	OpenAI	$0.000810	$24.30
GPT-5.6 Luna	OpenAI	$0.001320	$39.60

Compare every model side by side or price your own workload.

Budget contracts

Add a spendlint.yaml to your repository, describe the workload, and the GitHub Action reports PASS, WARN or FAIL with the projected before and after cost and the rule that broke.

version: 1
rules:
  max_monthly_usd: 500
  max_cost_per_request_usd: 0.02
  max_regression_pct: 20

npx spendlint check --contract spendlint.yaml --workload workload.json --baseline workload.baseline.json

For the swap above the command exits 1 and prints FAIL, the delta of +$540.00 per month (+150.00%), the violated rules and the cheaper compliant models. Check a contract in the browser, or read the setup guide.

What SpendLint does not do

It does not read your provider invoices or API keys. Every number is a projection from list prices and your declared workload.
It does not track spend after the fact. It stops the regression in review, where it is cheap to fix.
It does not guess prices. Every row carries the official source URL and the date it was checked, and stale rows raise a warning instead of quietly staying live.

Reference pages

Claude API pricing: Every Claude model with input, output, cache and batch prices.
OpenAI API pricing: GPT-5.x, o-series and mini models, cached input and batch.
Gemini API pricing: Gemini 3.x and 2.5 with long-context tiers and promotional periods.
Azure OpenAI pricing: Global Standard list prices from the Azure Retail Prices API.
Claude Code cost: What a developer seat costs per day and month, with Anthropic's own figures.
JSON API: The same dataset as machine-readable JSON with provenance on every row.