How LLM Live works

Every verdict on this site comes from rules you can read here. No vendor pays for placement, and the numbers below are the same ones the code uses.

Status

One verdict from four signals. The worst one wins.

Our checks: Where we hold an API key, a worker sends a tiny streaming request (count from 1 to 20, at most 80 tokens) to the vendor's own API every 60–120 seconds and records time to first token and tokens per second. A failed check is retried twice, 20 seconds apart. Rate limits and errors in our own setup are not counted. Each service page says whether our checks are running for it.
Rules: Last 3 checks failed → Major outage. At least half of the checks in 10 minutes failed → Partial outage. At least 20% failed, or the median time to first token is above 3 s and 3× slower than its 24-hour median → Degraded.
Official status pages: Polled every minute. A status older than 15 minutes is ignored, so a stuck page can't keep a service green or red.
Via OpenRouter: OpenRouter publishes the share of successful requests for every provider of every model. If at least 2 models on the vendor's own endpoints, and at least half of them, are below 80% over 30 minutes and 10 points below their 24-hour level → Degraded; below 50% on average → Partial outage. If 3 or more vendors fail at once, the problem is OpenRouter's and the signal is ignored.
User reports: At least 10 reports in 15 minutes and at least 4× the usual rate → Degraded. Reports alone never mark a service as down.
Recovery and alerts: The state is re-evaluated every 30 seconds. A service goes back to Operational only after 3 minutes without problems, so a flapping API doesn't spam alerts. Alerts go out when a problem starts, gets worse and ends.
Bars: Each bar is 30 minutes, coloured by the worst state in it. Grey means no data. Uptime is the share of bars without problems.

Models and prices

Every price comes from a public API and is checked every 5 minutes.

Source: The model catalog, prices and context windows come from OpenRouter's public API. Prices are per 1M tokens as OpenRouter lists them, which usually matches the vendor's own price but not always.
Price history: We keep a snapshot every 5 minutes and record every change, so price cuts and new or retired models show up in the news feed within minutes.
Where you can run it: The number of distinct providers serving a model on OpenRouter. Flex, batch, priority and spot tiers are not counted, because they are often unavailable by design.

Chess arena

Models play each other. A chess engine keeps the score honest.

A move: The model gets the position (FEN) and the move list and answers with a move and a short comment. Legality is checked with chess.js. After an illegal move the model gets the list of legal moves and tries again; the third illegal move loses the game. Three replies in a row without a move lose on time.
Thinking budget: Reasoning is capped at 1024 tokens per move so a game costs cents. This understates the strength of reasoning models, so the arena compares models under the same tight budget, not at their best.
Rating: Glicko-2, which accounts for how uncertain a rating is. The leaderboard is sorted by rating minus two deviations, a conservative estimate. A model is provisional until it has played 10 games. Pairings favour models with fewer games and close ratings.
Accuracy: Stockfish evaluates every position at depth 14. A drop in win chance of 5, 10 or 15 points is an inaccuracy, a mistake or a blunder; accuracy uses the Lichess formula.
Unrated games: A game that stops because of provider API errors is not rated. A game can also be voided by hand, with a published reason.

News

Our own data first, then the vendors' own words.

Our events: New models, price changes, retirements, incidents and arena leaders are generated from the data above, with no editing.
Outside sources: Vendor blogs via RSS, GitHub releases of major tools and Hacker News stories about AI with at least 150 points, checked every 15 minutes.
Editing: A small language model writes a headline and a one-line summary, rates importance from 1 to 5 and marks duplicates. It only sees the title and the RSS teaser, never the full article, and every item links to the source.

Limits

What this site can't see yet.

One location: Checks run from one server in Finland. A regional outage elsewhere may not show up, and a network problem on our side can look like a vendor problem; that's why one signal never decides alone when the others disagree.
Not every API: Some vendors have no machine-readable status page, and our own checks need an API key per vendor. Where neither exists, the verdict rests on OpenRouter data and user reports.
Open data: The current status of every service is available as JSON at /api/v1/status, free and without a key.