I tracked 400+ LLM API prices every day for a month. Here's what actually changed.

zoe wu — Mon, 14 Sep 2026 08:09:38 +0000

LLM pricing pages on the web are almost always stale. Providers reprice, deprecate, and launch models weekly, and most comparison posts are frozen the day they're published.

I got tired of it, so I built a tracker: every day at 03:17 UTC, a GitHub Action pulls the full OpenRouter catalog, diffs it against the previous day, and commits any price changes to a ledger. The result is LLMBill's price history — 445 models tracked, ~380 currently listed, every repricing event since July 21, 2026.

A month of data turned out to be more interesting than I expected.

August in numbers

251 repricing events in August alone (across ~380 live models)
129 price drops vs 99 hikes (≥1% threshold, to filter rounding noise)
Most active vendors: DeepSeek (54 events) and Qwen (54), followed by Z.ai (27) and Moonshot (18). OpenAI repriced only 13 times — frontier prices are sticky; the budget tier is where the knife fight is.

The biggest drops

Model	Input $/1M, Aug 1 → now	Net
Ling 3.0 Flash	$0.075 → $0.021	−72%
Nemotron 3 Super	$0.30 → $0.085	−72%
DeepSeek V4 Flash (0731)	$0.14 → $0.045	−68%
GPT-5.6 Sol (batch)	$2.50 → $1.00	−60%

The pattern: batch variants and budget/efficiency models get repriced aggressively. If your workload tolerates batch latency, August alone would have cut some bills by more than half.

The hikes nobody posts about

Price drops get launch posts. Hikes happen silently:

Model	Input $/1M	Net
Llama 3.3 70B Instruct	$0.13 → $0.91	+610%
Qwen3.6 27B	+108%
Gemini 3.7 Flash	+100% (doubled)

That +610% is mostly the free/discounted era ending for a legacy model — but if your pipeline hard-coded last year's price, that's a real incident waiting to happen.

What I took away as a builder

Budget-tier prices are volatile; frontier prices are not. GPT-5-class and Claude Opus-class prices barely moved all month. The 10–50× cheaper tier is where you win — and where you must watch for silent hikes.
"Set and forget" model routing is a bug. A quarterly review of your model choice is now a cost-control measure, not a nice-to-have.
Cached input pricing changes the math entirely (up to ~90% off repeated prefixes) — and cache prices reprice too.

How it's built

The whole thing is boring technology, deliberately: a 40-line Node script runs in GitHub Actions daily, diffs the OpenRouter API against the committed ledger, and pushes. Vercel rebuilds the static site on push, so the public numbers are never more than a day old. No database, no server, no cron box — git is the database.

The live boards: this week's price changes, plus a cost calculator if you want to price your own usage across all ~380 models.