ai-monitor: comparing AI provider prices based on your OpenCode sessions

⚡ The gist in 5 seconds

• A breakdown and repository called ai-monitor — monitoring of LLM provider prices and terms — has been published on Habr.
• You'll need Linux, Python 3.9+, OpenCode v2, Tavily and Artificial Analysis API keys, and a provider list in providers.yml.
• No quality testing of its own: the model ranking comes from the Artificial Analysis Intelligence Index, and the recommendation text is written by an LLM.

🔍 What happened

The Claude block for Russia and GitHub halting payments sent the author of the Habr post, who goes by Suntechnic, on a month-and-a-half run across providers. First he set up a FreeLLMAPI server — the free limits turned out to be tiny even combined. Topping up OpenRouter with $10 raised the daily limits on free models "by dozens of times," but most of those models had to be switched off: too small a context. OpenCode for $10 and even $40 didn't help either — the weekly limit burns out fast. After a week of searching, he had six paid providers on his hands, and the prices didn't differ by a single common multiplier: everyone had a different input-to-output token ratio, and some sold them at one price. It was out of this confusion that ai-monitor grew.

Here's how it works. The provider list lives in providers.yml, a file created based on the providers.example.yml template. OpenCode sessions are tagged by task type — via the opencode-session-tags plugin or the session_stats.py tag command. Then python3 session_stats.py collect gathers statistics, the author manually transfers the needed profiles from runs/session-stats/summary.yml into profiles.yml, and only then runs python3 ai_monitor.py. Search goes through Tavily (the free 1,000 requests per month are enough), and the model ranking comes through the Artificial Analysis API. The output is four Markdown files: recommendations, price comparison, model ranking, and sources, with the fresh set in reports/actual/. The author separately warns that when signing up for Tavily, he was being offered "1,500 free requests" that were actually on a paid plan.

🧠 Our take

The LLM access market has fragmented to the point where comparing providers by eye is meaningless. Subscriptions sell limits, not tokens, each has its own cache and context policy, and the same model can differ in price in different proportions for input and output across two sellers. Our assessment: demand for such scripts grew not because of technology, but because of access — blocks and payment issues pushed people into manual trial and error.

What does this change? The author honestly states the idea: not a universal list of "best models," but a calculation for your tasks and your spending profile — including cached tokens. For those already using OpenCode and keeping sessions, this saves an evening of arithmetic. For everyone else, it's a reason to develop the habit of tagging requests by task type, otherwise the profiles will be built on someone else's numbers.

Who wins is clear: the author himself and people with a similar Linux stack. Who loses: those expecting a ready-made answer of "go with this provider." There isn't one here and there won't be: quality is taken from someone else's index, and the recommendations are written by an LLM. We think that's a reasonable compromise for a raw personal tool, but not a replacement for your own measurements.

⚠️ What we don't know

The author doesn't provide a table of prices and limits for his six providers, doesn't name the savings amount, and doesn't show how well the recommendations matched reality. The model ranking is someone else's index, not a programming-specific benchmark. The advice text is generated by an LLM, so prices and terms should be double-checked manually before paying — Suntechnic himself says so. And yes, the README of his provider list contains referral links: the tool is free, but the author does have an incentive to feature specific sellers.

Prediction: within a quarter, the repository will get at least one measurement of its own (latency or cost for a specific task), otherwise the tool will remain a calculator of other people's prices.

Disclaimer: a single source — the author's post on Habr; the limit and savings figures are not officially confirmed.

Source: Habr