What Developers Are Actually Installing in May 2026
Stars on GitHub measure interest; package downloads measure use. This page pulls live monthly download counts from PyPI Stats and the npm registry download API for the major LLM vendor SDKs and the prominent multi-vendor abstraction layers. The picture is sharply different from the GitHub-star ranking, because download counts are a flow metric that updates monthly with actual deployment activity.
PyPI Monthly Downloads (Past 30 Days, May 14, 2026)
| Package | Vendor / Source | Monthly Downloads | Weekly Downloads |
|---|---|---|---|
| litellm | BerriAI (multi-vendor proxy) | 417,813,970 | 114,389,012 |
| openai | OpenAI | 295,890,968 | 71,743,491 |
| langchain | LangChain (framework) | 241,178,724 | 58,699,944 |
| google-genai | 216,729,048 | (rate-limited) | |
| anthropic | Anthropic | 119,776,565 | 28,829,540 |
| langchain-openai | LangChain (OpenAI binding) | 68,382,273 | 16,678,913 |
| cohere | Cohere | 41,094,289 | 10,365,083 |
| langchain-anthropic | LangChain (Anthropic binding) | 17,798,328 | 4,257,127 |
| groq | Groq | 17,438,817 | 4,036,538 |
| ollama | Ollama (local-LLM client) | 12,014,461 | 3,112,588 |
| xai-sdk | xAI (Grok) | 8,368,848 | 2,230,344 |
| replicate | Replicate | 1,351,206 | 291,015 |
Mistral (mistralai), Google's legacy google-generativeai, and Together AI's python SDK were rate-limited during this collection window and are omitted from the PyPI table.
npm Monthly Downloads (Past 30 Days, May 14, 2026)
| Package | Vendor / Source | Monthly Downloads |
|---|---|---|
| openai | OpenAI | 84,426,658 |
| @anthropic-ai/sdk | Anthropic | 71,587,111 |
| ai | Vercel AI SDK (toolkit) | 52,671,522 |
| @google/genai | 45,811,559 | |
| @ai-sdk/openai | Vercel AI SDK (OpenAI provider) | 25,552,001 |
| @ai-sdk/anthropic | Vercel AI SDK (Anthropic provider) | 23,122,504 |
| @langchain/core | LangChain (JS core) | 17,194,686 |
| @ai-sdk/google | Vercel AI SDK (Google provider) | 16,840,727 |
| @google/generative-ai | Google (legacy SDK) | 12,223,790 |
| @mistralai/mistralai | Mistral | 11,311,551 |
| @langchain/anthropic | LangChain (Anthropic binding) | 3,221,928 |
| groq-sdk | Groq | 2,999,759 |
| ollama | Ollama JS client | 2,183,406 |
| replicate | Replicate | 1,905,228 |
| cohere-ai | Cohere | 1,786,801 |
Eight Things the Downloads Tell You
- LiteLLM downloads exceed OpenAI Python SDK downloads by ~40 percent. 418M vs 296M monthly. Multi-vendor routing has gone mainstream; most production agent stacks now sit behind a vendor abstraction layer rather than calling OpenAI directly.
- Google's rebranded SDK (google-genai) at 217M/month overtook the Anthropic Python SDK. The migration from the legacy google-generativeai package to google-genai is mostly complete, and Google's Python adoption is now solidly in second place behind OpenAI on raw downloads.
- OpenAI vs Anthropic gap is much tighter on npm than on PyPI. PyPI: 296M vs 120M (a ~2.5x gap). npm: 84M vs 72M (a ~1.2x gap). JavaScript/TypeScript developers are far more vendor-balanced than Python developers, possibly reflecting where the early-2026 agent applications are being built.
- Vercel AI SDK provider packages add up to 66M/month. @ai-sdk/openai (26M) + @ai-sdk/anthropic (23M) + @ai-sdk/google (17M) = 66M scoped-package installs, on top of the 53M downloads of the toolkit-level ai package. The Vercel SDK is the dominant abstraction layer in the JavaScript ecosystem; the equivalent role on Python is split between LiteLLM and LangChain.
- LangChain still dominates Python AI tooling. langchain itself does 241M/month, langchain-openai another 68M, langchain-anthropic 18M. Including the JS @langchain/core at 17M, the LangChain ecosystem moves roughly 345M monthly downloads, a comparable footprint to OpenAI direct.
- Cohere is Python-heavy. Cohere PyPI: 41M/month. Cohere npm: 1.8M/month. The 23x Python:JS ratio reflects Cohere's enterprise-RAG positioning, which skews Python. By comparison, OpenAI Python:JS is roughly 3.5x.
- xAI is small but real. xai-sdk at 8.4M/month on PyPI is roughly 5 percent of Anthropic's footprint and 3 percent of OpenAI's, but it is materially larger than Replicate, Cohere npm, or Ollama JS. Grok has a developer ecosystem; it is just smaller than the leading three.
- Ollama JS lags Ollama Python by 5.5x. 12M PyPI vs 2.2M npm. Local-LLM deployment is still heavily Python-coded.
Adoption Rankings (Combined Python + npm Estimate)
For brand-visibility teams who care about "which vendor SDK do developers actually use," here is a combined Python + npm direct-SDK ranking in May 2026:
| Rank | Vendor | Combined Monthly Downloads (M) |
|---|---|---|
| 1 | OpenAI | ~380 |
| 2 | ~275 | |
| 3 | Anthropic | ~191 |
| 4 | Cohere | ~43 |
| 5 | Groq | ~20 |
| 6 | Mistral | ~11 |
| 7 | xAI | ~8 |
| 8 | Replicate | ~3 |
Mistral and Replicate Python figures excluded (collection rate-limited); the Mistral total is conservative.
What This Means for AI Visibility
SDK downloads predict which vendor APIs are most often inside the recommendation pipeline that surfaces a brand to an end user. The 4-to-1 ratio between OpenAI Python downloads and the next-largest Anthropic Python downloads means that most production agent stacks talk to OpenAI by default. But the LiteLLM and Vercel AI SDK numbers suggest a structural shift: a growing share of agent traffic is vendor-abstracted at install time, which means the developer chose multi-provider routing rather than a single vendor. For brands optimising visibility, the implication is that single-platform tracking (ChatGPT-only or Claude-only) increasingly underestimates the actual surface where recommendations happen because the deployment may swap providers per call.
Methodology
Monthly download counts pulled from the PyPI Stats public API and the npm registry download endpoint on May 14, 2026. PyPI Stats reports a 30-day rolling window; npm reports calendar-bounded last-month. Both are subject to bot-installation noise (CI/CD pipelines, mirroring caches), which inflates absolute counts but does not meaningfully change relative rankings between similar-shape packages. A small number of Python packages (google-generativeai legacy, mistralai, together) were rate-limited at collection time and are noted as omitted in the source tables. Refreshed quarterly. Treat figures as accurate at time of capture.
How Presenc AI Helps
Presenc AI tracks brand-mention rates across the major AI platforms whose SDKs are ranked above. When a brand's mention rate on OpenAI rises while its mention rate on the abstracted multi-vendor stack (LiteLLM, Vercel AI SDK) stays flat, that gap is the signal that the brand is winning on the consumer-facing platform but losing on the developer-facing distribution. SDK adoption data closes the loop between "who runs my brand through which model" and "does my brand surface in the answer."