About

• OpenAI models.md

  Catalog only — the page prints no prices; they come from openai-pricing-md. The marketing page (openai.com/api/pricing) is Cloudflare-403'd to any headless client and must never be a dependency.

• OpenAI pricing.md

  Did not exist at the 2026-08-25 audit; live-checked 2026-09-22 as real markdown (text/markdown, ETag, Last-Modified) at the docs origin, not the 403'd marketing page. One row per model, service tier and context band; the Standard short-context row shares its key with the catalog row from openai-models-md. The "(<272K context length)" annotation is the short-context bound and is kept in notes, never read as a context window. Cache-write prices, the cyber / multimodal / realtime "Grouped Pricing Table" sections and per-minute or per-image tables are not $/MTok token pricing and are not parsed. Promotional windows are prose on this page ("promotional pricing is available at least through …") with no hard end date, so no effective_until is set — the price shown is the price printed.

• Anthropic models overview.md

  Catalog fields (context window, cutoff); pair with anthropic-pricing-md for prices.

• Anthropic pricing.md

  Audit recorded this page as semantic HTML; scaffold-time fetch (2026-08-25) confirmed the .md suffix serves real markdown here too.

• xAI models.md

  Long-context dual pricing: models listed with two rows (< 200k / >= 200k prompt tokens). The marketing page (x.ai/api) is Cloudflare-403'd and must never be a dependency.

• Google Gemini API pricing

  Model names live in headings above table clusters; prices carry dual-date promo strings ("$0.75 through Dec 31, 2026. $1.50 starting Jan 1, 2027."). The .md trick returns a JS shell and the models page is client-rendered — only this pricing page is genuine SSR.

• OpenRouter /api/v1/models

  Third-party restatement with no drift SLA and thin xAI coverage (6 vs 20+ SKUs) — cross-check, never source of record.

• OpenRouter usage rankings (daily)

  OpenRouter share, not market share — one aggregator's traffic. Direct API and enterprise volume never route through OpenRouter, so a lab's share here measures its standing among developers who use an aggregator, not its revenue. Fetched without date parameters: the endpoint's default window is the trailing 30 completed UTC days, so every 6-hourly survey tick re-reads the same days and a late revision lands as a modified row. Rate limits 30 req/min per key, 500/day per account; this terminal uses four a day.

• OpenAI Ashby job board

  757 roles at audit time.

• Anthropic Greenhouse jobs

  537 roles at audit time. Department requires the /departments join — the jobs endpoint carries no department field.

• xAI Greenhouse jobs

  262 roles at audit time; 255 at the 2026-08-25T23:08Z join re-fetch. Every row carries company_name "SpaceXAI" — the board is an entity-blended SpaceX/xAI artifact of the 2025 merger, and presenting it as pure xAI hiring would mislead. Caveat must ship visibly wherever this data surfaces. What the /departments join adds is bounded evidence, not a fix: the 26 departments are xAI-shaped (Data Center 121, Human Data 36, Engineering 23) with no aerospace or launch function, and 2 of 255 titles name a SpaceX-side program (Starlink, battery storage). Department requires the join — the jobs endpoint carries no department field.

• xAI Greenhouse departments (join)

  Departments are the board owner's own labels, not an independent entity split: they narrow how much SpaceX could be hiding in the blend, they do not prove none is.

• EDGAR full-text search (S-1 tripwire, “Anthropic”)

  UA header required (contact email form), 10 req/s. Open keyword sweeps across all filers are confirmed noisy (micro-cap false positives) — whitelist only. Confidential DRS filings are structurally invisible until flipped public; we catch the public-flip moment and say so.

• EDGAR full-text search (S-1 tripwire, “OpenAI”)

  The OpenAI side of the S-1 tripwire, same UA + rate rules and the same noise as edgar-fts: an S-1 that mentions OpenAI in its text is a hit, so display_names (the filer) is what the resolver matches, never the query term. The endpoint caps a response at 100 hits.

• EDGAR submissions (SPCX — xAI’s parent)

  Same UA + rate rules as edgar-fts. The parser keeps only the tracked forms (S-1, 424B4, 10-Q, 10-K, 8-K and their amendments — server/pipeline/parsers/sec/forms.ts); Form 4, 144, 3, 13G and D rows are dropped at parse so insider trades and notices never reach the judge as "an SEC change". SPCX is xAI's parent since the 2025 merger: its periodic filings are where any xAI segment disclosure would land.

• EDGAR submissions (MSFT)

  Whitelisted CIK, same form filter as edgar-submissions. Microsoft's 10-Q/10-K carry the OpenAI equity-method and commitment disclosures; the terminal alerts on the filing, it does not extract the figure.

• EDGAR submissions (AMZN)

  Whitelisted CIK, same form filter as edgar-submissions. Amazon's 10-Q/10-K carry the Anthropic convertible-note and investment disclosures; the terminal alerts on the filing, it does not extract the figure.

• EDGAR submissions (NVDA)

  Whitelisted CIK, same form filter as edgar-submissions.

• EDGAR submissions (GOOGL)

  Whitelisted CIK, same form filter as edgar-submissions. Alphabet is Google's parent; its filings segment Google Cloud, never Gemini, so nothing here is a lab revenue figure (see revenue_filers).

• EDGAR XBRL company facts (SPCX — xAI’s parent)

  SpaceX consolidated revenue, not xAI's: SpaceX is xAI's parent since the 2025 merger and files as one registrant. Company facts carry non-dimensional XBRL facts only — no segment members — so whether SpaceX segments xAI in its 10-Q cannot be read here and is unverified. Read as the parent's audited top line, the ceiling any xAI figure sits under, never as a lab revenue number. Revenue tags are read in a fixed order (Revenues, RevenueFromContractWithCustomerExcludingAssessedTax, SalesRevenueNet) and every tag present is stored with its name; the reader picks per period. Same UA + rate rules as edgar-fts.

• OpenAI status page incidents

  Capacity-strain proxy, not an SLA: a status page records what the vendor chose to post. Live-fetched 2026-09-08: 25 incidents (2026-08-05 to 2026-09-08), so the feed itself covers roughly one month and older history is only what this pipeline accumulates (append-only lane — an incident scrolling out of the feed is kept). No components on this page; the incident URL is composed from the feed's own page.url + /incidents/{id} (verified to resolve), not read from a field.

• Anthropic (Claude) status page incidents

  Registered at the final URL: status.anthropic.com 301-redirects to status.claude.com. Live-fetched 2026-09-08: exactly 50 incidents (2026-07-21 to 2026-09-03), which is the page's window cap — counts before that date are partial until the pipeline has polled through it. Components name the surface (Claude API, claude.ai, Claude Code, Claude Cowork, Console); consumer-app incidents count alongside API ones, so read the component list, not just the count. Incident URL composed as page.url + /incidents/{id} (verified to resolve).

• Google Cloud status incidents (Gemini / Vertex AI only)

  This is the Alphabet-wide Cloud feed; an unfiltered count would conflate Compute/BigQuery/Spanner outages with Gemini and mislead — the same reasoning that cuts Google hiring. The parser keeps ONLY incidents whose affected_products[].title contains "Gemini" (case-insensitive) or equals "Vertex AI"; everything else is dropped and counted in the run note. Live-fetched 2026-09-08: 6 incidents across 40 product titles, 1 kept ("Vertex Gemini API", 2026-02-27). Severity vocabulary is Google's (low/medium/high), not Statuspage's (minor/major/critical); status is the most recent update's status (AVAILABLE once resolved). Incident URL composed as status.cloud.google.com/ + uri (verified to resolve).

• google-hiring

  cut — display "no public feed" honestly. None usable: careers.google.com is an auth-walled SPA; the DeepMind Greenhouse board is vestigial (10 jobs). Alphabet-wide numbers would conflate Search/Ads/Cloud with Gemini and mislead.

• artificial-analysis

  cut — external link only, never ingested. artificialanalysis.ai/data-api (read 2026-09-08) states the free API tier is "Internal use only; no redistribution." and that redistribution rights come only with a commercial package ("Commercial redistribution with attribution"). A public terminal republishes by definition, so the site is linked as a reference beside the demand-share panel and no figure from it is stored or shown.

• lmarena

  cut — external link only, never ingested. arena.ai (formerly lmarena.ai; read 2026-09-08) publishes a leaderboard app with no documented data endpoint, dataset download, or data license. Scraping a page with no stated license would be a claim about that page, not data; it is linked as a reference beside the demand-share panel.

• compute-deals

  stretch goal only (new-post tier, no extraction claims). RSS detection solvable for 4 sources (Anthropic = SPA no feed; xAI = hard 403), but the deals that matter break in the press before any company feed — honest coverage is impossible, so extraction is cut.

• xai-status

  cut — display "no public feed" honestly. status.x.ai/api/v2/incidents.json answers 403 (Cloudflare challenge page) to any non-browser client, live-checked 2026-09-08 with a descriptive User-Agent. Scraping a rendered page would be a dependency on a bot wall; the terminal shows the gap instead.

• mistral

  cut — not tracked on any axis. No supported surface on either primary axis, live-checked 2026-09-08. Hiring: api.ashbyhq.com/posting-api/job-board/mistral answers 404 (the board at jobs.ashbyhq.com/mistral is a 7 KB HTML shell over an internal GraphQL endpoint, not the posting API). Pricing: mistral.ai/pricing is a 485 KB client-side HTML app (same bytes at docs.mistral.ai/deployment/laplateforme/pricing) with no markdown or JSON endpoint; docs.mistral.ai/llms.txt lists a Pricing and a Models Overview .md page, but both links answer 404 with a 462 KB HTML shell. An HTML parse in the Google style is conceivable; with no hiring feed to pair it with, a fifth lab on one axis is not worth the parser.

SEC filings

The EDGAR feeds — the S-1 full-text tripwire, one submissions feed per whitelisted CIK, and XBRL company facts for each tagged revenue filer — are polled every 30 minutes: the one axis where latency is worth polling for.

Pricing and hiring

Every source, EDGAR included, is surveyed every 6 hours. The overlap is deliberate: a stalled tripwire goes covered, not blind.

Pages

Each payload is cached for five minutes, keyed by the newest poll tick, so a poll invalidates the cache by construction. The tick a page reflects is printed on it.

Confidential filings

A confidential draft registration is structurally invisible on EDGAR until it flips public. The terminal catches the flip and says so; it cannot see the draft.

This is a data terminal, not a recommendation. It reports what public sources say and when they said it. Nothing here is personalised, nothing here is a solicitation, and the judged alerts are a model’s reading of change rows, labelled as such.

• Source on GitHub — MIT licensed; the source registry and every parser are in the repo.
• Bulk exports — every table as CSV or JSON.
• Atom feed — alerts, for a reader that wants to be told.