Sources

Anthropic's public status page (Statuspage-style). The machine-readable feed at /api/v2/status.json (redirects to status.claude.com) is consumed by the vendor-status observer at lib/observers/anthropic.ts. Vendor-reported; not an independent uptime monitor.

Probe target only — the bare api.anthropic.com host responds 404 to unauthenticated GETs at /, which tells us the host is reachable without invoking any inference endpoint. Hourly probe via /api/cron/status.

Google Cloud's public incidents feed (JSON array of incident records). The vendor-status observer filters to incidents whose affected_products mention Gemini / Vertex AI / AI Studio. Vendor-reported; not an independent uptime monitor.

Anthropic's models overview page (legacy section), including the Claude Opus 4 row.

Anthropic's pricing reference (model pricing, prompt caching multipliers, batch API pricing).

Anthropic's public status page (Statuspage-style). The machine-readable feed at /api/v2/status.json (redirects to status.claude.com) is consumed by the vendor-status observer at lib/observers/anthropic.ts. Vendor-reported; not an independent uptime monitor.

Probe target only — the bare api.anthropic.com host responds 404 to unauthenticated GETs at /, which tells us the host is reachable without invoking any inference endpoint. Hourly probe via /api/cron/status.

Per-model specification page for gemini-2.5-pro (model code, token limits, supported input types, knowledge cutoff, supported features).

Gemini API pricing reference — standard, batch, flex, and priority tiers with ≤200k vs >200k-token pricing.

Quickstart with REST endpoint shape, model parameter format, request body, and Python/JS SDK examples.

Google Cloud's public incidents feed (JSON array of incident records). The vendor-status observer filters to incidents whose affected_products mention Gemini / Vertex AI / AI Studio. Vendor-reported; not an independent uptime monitor.

Meta Llama 4 model card (Scout + Maverick). Context windows (Scout 10M, Maverick 1M), input modalities (text + up to 5 images), output modality (text only), parameter counts (17B active), knowledge cutoff (August 2024). Max output tokens and a stated release date are NOT on the page. Meta does not run a hosted paid API for Llama; pricing stays null.

Mistral Large 3 spec card at the new URL pattern (/models/model-cards/<slug>). Canonical pinned snapshot is mistral-large-2512; context window 256k; input $0.5 / output $1.5 per 1M tokens. Max output and explicit modality list still not stated.

Mistral models overview — current active models, version stamps, open vs premier classification.

Mistral models table — exact API string for every current model (mistral-large-3, mistral-medium-3-5, etc.).

Mistral chat completions endpoint, headers, minimal request body, and model parameter format.

DeepSeek API Models & Pricing page — current model IDs (deepseek-v4-flash, deepseek-v4-pro), context window, input cache-miss / cache-hit / output rates, time-boxed discount window.

DeepSeek chat completion API reference — endpoint, model parameter values (deepseek-v4-flash, deepseek-v4-pro), minimal request body.

DeepSeek API docs root — active vs deprecated model status (deepseek-chat and deepseek-reasoner marked for deprecation 2026/07/24).

Historical release note for DeepSeek-R1-0528 (May 28, 2025). Used to anchor the historical deepseek-r1 catalogue entry — R1 family is not in the current API model parameter set.

Groq's public pricing page. Confirms Llama 4 Scout (17Bx16E) 128k at $0.11/1M input, $0.34/1M output. Page also lists prompt caching (50% discount on cache-hit input) and a Batch API (50% off, 24h–7d processing window). Hosted-provider context — Groq is the billing provider, Meta is the model creator.

Together AI's public pricing page. Confirms DeepSeek V4 Pro on serverless at $2.10/1M input, $4.40/1M output, with a separately listed $0.20/1M cache-hit input rate. Llama 4 Scout and Llama 4 Maverick appear only in the Fine-Tuning table (LoRA / Full Fine-Tuning columns), not in the serverless inference table — Sprint 19 does NOT record those numbers as inference pricing.