frontierterm.com

Orientation: the site’s llms.txt — what it tracks, the provenance rule every number obeys, and the pages behind each tool.

The signal band.

The API price catalog: one row per SKU key (provider + model + tier) at its newest revision, in USD per million tokens (input_per_mtok, cached_input_per_mtok, output_per_mtok; null = the vendor publishes no figure), plus delta vs.

Open roles per lab, counted from the CURRENT job-board set (a role missing from the board is closed), by department.

Alerts newest first, each citing the change rows it is a claim about (changes, resolved; missing_change_ids if any did not resolve).

One alert by id (ids come from get_alerts, get_overview, the Atom feed or an /alerts/<id> permalink), with every cited change row in full: fields lists every scalar field before and after, unchanged ones included.

Every observation of one SKU key, oldest first: points carry kind (baseline = the price as of the source’s first snapshot, added, modified with the changed fields in diff, removed, or current), the three USD-per-million-token figures, and the fetch each was read from.

Open roles per lab and department, per UTC day (dates, total, departments[].series), RECONSTRUCTED by undoing the change log backwards from the current job-board set — not a series of daily snapshots.

Every SKU first listed on a vendor page after that source’s baseline, newest first, with the price printed beside it when it was first seen (first_seen_at is the detection, not the vendor’s launch date) and also_listed_by for keys two pages carry.

Capacity strain as the vendor chose to post it: three status feeds (OpenAI and Anthropic Statuspage, Google Cloud filtered to Gemini / Vertex AI products).

Each lab’s own index of misalignment reports and notices — agents leaving their sandbox, messaging each other through public services, deceiving in their own summaries.

OpenRouter’s daily usage rankings folded into token share per lab over the newest 7 days present (shares[].share, 0..1) with the change against the 7 before (delta_pp, percentage points; null when there is no prior window), each lab’s top models in the window, and a daily series.

Revenue each lab, or the public parent that consolidates it, has reported to the SEC, read from EDGAR’s XBRL company facts.

The 7-day rankings joined to the vendors’ list prices: per lab, tokens routed, the share of those tokens that matched a priced SKU (coverage), and implied spend as a RANGE — spend_low is every token at the input list price, spend_high every token at the output price (the rankings carry no prompt/completion split, so a point estimate would be invented).

Every registered source: its audited URL, role (primary, join, cross-check, sec), the audit’s caveat in its own words, newest snapshot (with HTTP status and byte count), how the last poll ended (ok, unchanged, failed, baseline), snapshot count and how many current rows came from it.

Is what is running the thing that was meant to be running: build sha, migrations the repo has vs.