| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| tags: |
| - llm |
| - latency |
| - benchmark |
| - inference |
| - observability |
| - api-monitoring |
| pretty_name: "LLM Latency Tracker: latency and uptime for AI inference APIs" |
| size_categories: |
| - 1K<n<10K |
| configs: |
| - config_name: daily_aggregates |
| data_files: daily_aggregates.csv |
| default: true |
| --- |
| |
| # LLM Latency Tracker |
|
|
| Independent, continuously measured **latency and availability for AI inference API |
| providers**, aggregated by day. Covers **45 providers** across |
| **4 regions** (ap-tokyo, eu-hetzner, sa-east, us-central), built from |
| **1,347,917 raw probes** collected between 2026-07-23 and |
| 2026-08-18. |
|
|
| Live rankings and full methodology: **[llmlatency.dev](https://llmlatency.dev)** |
|
|
| ## How the numbers are produced |
|
|
| Probes run every five minutes from separate network locations and are never routed |
| through a gateway or an aggregator, so the numbers describe the providers themselves |
| rather than a proxy in front of them. |
|
|
| - **Network probe** — DNS → TCP → TLS → time to first byte (`ttfb`). |
| - **Inference probe** — time to first token on a real completion request (`ttft`). |
|
|
| These are different quantities, an order of magnitude apart, and they are never mixed in |
| one ranking. Percentiles are nearest-rank, identical to the ones shown on the site. |
|
|
| ## Files |
|
|
| | file | contents | |
| |---|---| |
| | `daily_aggregates.csv` | one row per date × provider × region × probe type, with p50/p95, sample count and success rate | |
| | `rankings.json` | machine-readable snapshot published live at llmlatency.dev | |
|
|
| ## Usage |
|
|
| ```python |
| from datasets import load_dataset |
| |
| ds = load_dataset("llmlatency/llm-latency-tracker", split="train") |
| ds.filter(lambda r: r["region"] == "eu-hetzner" and r["probe_type"] == "network") |
| ``` |
|
|
| Or without this dataset at all — there is a client on PyPI: |
|
|
| ```bash |
| pip install llmlatency |
| llmlatency fastest |
| ``` |
|
|
| ## Limitations, stated up front |
|
|
| Vantage points are cloud data centres, not consumer networks, so absolute values are |
| lower than an end user would see — the comparison between providers is the meaningful |
| part. Provider coverage changes over time as APIs appear and shut down. |
|
|
| ## Citation |
|
|
| Archived, citable version with a DOI: <https://doi.org/10.5281/zenodo.21954788> |
|
|
| ```bibtex |
| @dataset{llmlatency_tracker, |
| title = {LLM Latency Tracker: measured latency and uptime for AI inference APIs}, |
| author = {llmlatency.dev}, |
| year = {2026}, |
| doi = {10.5281/zenodo.21954788}, |
| url = {https://llmlatency.dev} |
| } |
| ``` |
|
|
| ## Links |
|
|
| - Live site and methodology: <https://llmlatency.dev> |
| - Machine-readable API: <https://llmlatency.dev/api/rankings.json> |
| - Daily snapshots in git: <https://github.com/mazamaka/llm-latency-tracker> |
| - Python client: <https://pypi.org/project/llmlatency/> |
| - Citable archive (DOI): <https://doi.org/10.5281/zenodo.21954788> |
|
|
| Licence: **CC-BY-4.0**. Snapshot generated 2026-08-18. |
|
|
