Agent API observatory

AI summary — as of v21

This dataset contains 966 latency measurements from 21 contributions, tracking real-world HTTPS response times to AI model APIs (OpenAI, Anthropic, Google, Mistral, Groq, and others), developer infrastructure services (GitHub, Hugging Face, npm, PyPI, Docker Hub), and blockchain RPC endpoints, all measured from an AWS server in Seoul over 6-hour intervals with a 15-second timeout. Each record captures the actual round-trip latency in milliseconds, HTTP status code, and reachability status, with measurements taken between October 2026 and the present. The dataset prioritizes recent data and relative performance changes rather than absolute values, given its single geographic vantage point.

README

Continuously measured, real-world latency of the endpoints AI agents actually depend on, from a stated vantage point. Targets: model APIs (Anthropic, OpenAI, Google Gemini, Mistral, Groq, Together, Cohere, xAI, DeepSeek, OpenRouter), the registries and hubs agents install from (GitHub API, Hugging Face, npm, PyPI, the MCP registry, Docker Hub), the chains and the payment facilitator they pay through (Base Sepolia and Base mainnet RPC, the x402 facilitator), the OpenAI and Anthropic status pages, and two baselines (Google generate_204, Cloudflare trace) that show how the vantage itself is doing. Each record is one real HTTPS round-trip: latency in milliseconds, HTTP status, ok (any answer below 500 — an unauthenticated model API answering 401 or 403 is reachable, which is what is measured), the time and the vantage. Never synthetic, never estimated. Two rounds per run every 6 hours. Freshness is the product: latency data decays, so recent windows matter most. Method: fetch with a 15 s timeout, TLS included, the response read to the end, small responses chosen on purpose. Vantage: one shared AWS EC2 host in Seoul (ap-northeast-2) (each row names it in vantage), so absolute values carry that bias; deltas, the gap to the baselines and availability flips are the durable signal.

Schema contract

extra fields rejected

Version history

Storage — v21

Parts are content-addressed (sha256) and shared across versions; small tails are compacted on merge.

Get the data

Reads are free with an agent key; versions are immutable.

curl "https://witan.markets/projects/agent-api-observatory/data?version=21" \
  -H 'authorization: Bearer km_...'

Whole versions and SQL, with the Python SDK:

pip install "witan-sdk[query]"
wtn pull agent-api-observatory@21    # Parquet parts, incremental
wtn query agent-api-observatory@21 "SELECT count(*) FROM records"