Which AI provider SDKs are actually installed
· refreshed weekly
anthropic has the most-installed first-party AI API SDK at 316,942,230 downloads over the last 30 days, across PyPI and npm. Installs are not latency and not quality — they are the other question a developer asks before picking an API: is anyone actually using this.
Most-installed AI provider SDKs
| # | Provider | Total | PyPI | npm |
|---|---|---|---|---|
| 1 | anthropic | 316,942,230 | 175,863,980 | 141,078,250 |
| 2 | 304,892,321 | 218,957,289 | 85,935,032 | |
| 3 | mistral | 60,708,342 | 28,830,905 | 31,877,437 |
| 4 | cohere | 24,965,181 | 23,092,470 | 1,872,711 |
| 5 | groq | 22,401,123 | 18,836,766 | 3,564,357 |
| 6 | voyage | 4,510,591 | 3,608,376 | 902,215 |
| 7 | xai | 3,766,973 | 3,766,973 | 0 |
| 8 | replicate | 3,414,719 | 1,415,681 | 1,999,038 |
| 9 | qwen | 2,604,108 | 2,604,108 | 0 |
| 10 | together | 2,001,621 | 1,669,414 | 332,207 |
| 11 | fireworks | 1,943,557 | 1,943,546 | 11 |
| 12 | cerebras | 1,579,965 | 1,046,105 | 533,860 |
| 13 | glm | 280,580 | 280,580 | 0 |
| 14 | ai21 | 263,420 | 263,073 | 347 |
| 15 | writer | 247,925 | 225,995 | 21,930 |
| 16 | reka | 112,382 | 112,382 | 0 |
| 17 | baseten | 2,385 | 2,385 | 0 |
| 18 | minimax | 1,838 | 1,838 | 0 |
Who grew and who shrank in the last 30 days?
| Provider | Was | Now | Change |
|---|---|---|---|
| reka | 70,608 | 112,382 | +59.2% |
| ai21 | 243,381 | 263,420 | +8.2% |
| anthropic | 313,609,889 | 316,942,230 | +1.1% |
| voyage | 4,516,178 | 4,510,591 | -0.1% |
| replicate | 3,898,107 | 3,414,719 | -12.4% |
| cerebras | 1,848,193 | 1,579,965 | -14.5% |
| 360,201,090 | 304,892,321 | -15.4% | |
| xai | 4,478,838 | 3,766,973 | -15.9% |
| openai | 561,278,321 | 462,337,104 | -17.6% |
| qwen | 3,236,984 | 2,604,108 | -19.6% |
| mistral | 76,624,752 | 60,708,342 | -20.8% |
| baseten | 3,156 | 2,385 | -24.4% |
| groq | 29,927,789 | 22,401,123 | -25.1% |
| together | 2,695,651 | 2,001,621 | -25.7% |
| glm | 388,323 | 280,580 | -27.7% |
Why is the openai package not in this table?
Downloads of the `openai` package are not attributed to any provider: most OpenAI-compatible providers document their own API through it, so the number measures the protocol's reach, not one vendor's usage.
What do these numbers actually measure?
Package downloads from the official PyPI and npm statistics endpoints, over the last 30 days. A download is a machine fetching the package — CI runs, Docker builds and mirrors are all in there, so this is a measure of reach, not of active users, and certainly not of quality. It is published because it is verifiable: the same two endpoints are open to anyone who wants to check the figures.
Sources: https://pypistats.org/api · https://api.npmjs.org/downloads. Latency for these same providers is measured separately — see the latency rankings.