Chart of the Day · OCT 19, 2026

Anthropic SDK Passes OpenAI in npm Downloads

In the week of 28 Sep, @anthropic-ai/sdk was downloaded 53.7M times on npm, +6% against the openai package. A year earlier it trailed by 66%.

Mode VC ·

The signal

@anthropic-ai/sdk went from −66% to +6% against openai in weekly npm downloads between the weeks of 6 Oct 2025 and 28 Sep 2026.

Weekly npm downloads of each package as % above or below the openai package (0 = parity). Complete Mon to Sun weeks, 6 Oct 2025 to 4 Oct 2026.

Anthropic SDKGoogle GenAIVercel ai
Sources: npm registry download counts (api.npmjs.org) · TechCrunch, 16 Oct 2026: Vercel AI gateway adoptionMODE VC

<b>What this counts:</b> npm downloads include CI runs and installs pulled in as a dependency, so they track how often code that uses a package is built, not how many developers use it. Weeks with a missing day are left out (45 complete weeks remain). Anthropic was also ahead in 1 earlier week (25 May) before falling behind again, so two weeks in front is a lead, not yet a trend.

Show the data (45 rows)
Week ofopenai@anthropic-ai/sdkvs openai@google/genaiai (Vercel)
2025-10-067.3M2.5M−66%1.9M3.6M
2025-10-137.5M2.5M−66%2.2M3.7M
2025-10-207.4M2.7M−64%1.6M4.0M
2025-10-277.3M2.6M−64%1.6M4.2M
2025-11-037.7M2.7M−64%1.7M4.6M
2025-11-107.7M2.8M−63%1.8M4.8M
2025-11-178.1M3.1M−61%2.1M5.1M
2025-11-247.3M2.8M−62%2.0M4.9M
2025-12-018.8M3.4M−62%2.3M5.5M
2025-12-088.7M3.4M−61%2.4M5.2M
2025-12-158.5M3.3M−62%2.6M5.3M
2025-12-225.3M2.2M−58%2.0M3.1M
2025-12-295.3M2.2M−57%2.1M2.8M
2026-01-058.8M3.5M−60%3.3M5.6M
2026-01-129.6M3.7M−61%3.4M6.2M
2026-01-1910.4M4.1M−60%3.7M6.4M
2026-01-2611.9M5.2M−56%4.6M7.0M
2026-02-0212.2M5.7M−53%4.9M7.6M
2026-02-0912.4M5.8M−53%5.7M7.8M
2026-02-1611.9M5.8M−51%5.8M8.0M
2026-02-2313.8M6.9M−50%8.1M8.7M
2026-03-0214.9M7.7M−49%8.8M9.9M
2026-03-0916.8M8.8M−48%9.8M10.1M
2026-03-1616.4M9.1M−45%9.3M9.8M
2026-03-2316.4M10.4M−36%9.1M10.4M
2026-03-3015.7M12.4M−21%9.1M10.1M
2026-04-0617.1M13.4M−22%9.6M10.6M
2026-04-1317.9M14.8M−18%10.2M11.5M
2026-04-2019.3M15.4M−20%10.6M11.9M
2026-04-2721.1M18.4M−13%11.2M12.1M
2026-05-0420.5M17.1M−17%10.4M12.6M
2026-05-1121.6M17.9M−17%10.9M13.2M
2026-05-1822.5M19.5M−13%11.9M13.7M
2026-05-2524.0M24.4M+2%13.5M13.0M
2026-06-0825.8M23.5M−9%15.9M15.1M
2026-06-1527.4M24.4M−11%17.0M15.7M
2026-06-2925.4M23.1M−9%15.3M15.1M
2026-07-1329.5M27.1M−8%16.2M17.7M
2026-07-2030.6M27.6M−10%16.4M18.8M
2026-07-2732.7M29.7M−9%17.8M19.8M
2026-08-0332.0M29.8M−7%17.5M20.6M
2026-08-1734.5M33.9M−2%21.2M22.4M
2026-08-2439.7M38.7M−3%23.4M23.6M
2026-09-2144.9M46.4M+3%26.7M30.4M
2026-09-2850.9M53.7M+6%29.5M34.5M
Cite this dataset & embed chart
APA Citation:Mode VC Research. (2026). Anthropic SDK Passes OpenAI in npm Downloads [Data set]. Mode VC. https://modevc.com/charts/anthropic-sdk-npm-downloads-vs-openai/
Embed Snippet (HTML):<iframe src="https://modevc.com/charts/anthropic-sdk-npm-downloads-vs-openai/" width="100%" height="600" frameborder="0" title="Anthropic SDK Passes OpenAI in npm Downloads"></iframe>

The Mode read

Developers pick a model when they write the import line. Swapping a model later means touching code, tests and prompts, so the SDK in the build is a stickier signal than a leaderboard rank. TechCrunch's 16 Oct report on Vercel's gateway showed how fast a new model can win teams that are already wired to switch. So which lab SDK are JavaScript builds pulling in?

For most of the year, the answer was openai. In the week of 6 Oct 2025, @anthropic-ai/sdk trailed it by 66%. In the week of 28 Sep it was +6% ahead: 53.7M downloads against 50.9M. Anthropic's package grew 21 times over the period, openai's 7 times and @google/genai's 15 times.

A lab SDK is a distribution channel: every agent, IDE plugin and internal tool that imports it ships the lab's model by default. The lead is two weeks old and npm counts builds, not people, so it can flip back. The useful lesson for founders is where defaults get set. If your product sits on top of one lab, make the model a setting, not an import, because the default is moving.

The model a developer imports is the model that ships. That default just moved.

What we would back: Developer tools that let teams switch model providers without rewriting code, and win the default slot in agent frameworks.

Every number is computed at build time from the cited primary sources. Machine-readable: JSON.