Research / Edition 2026-09
AI infrastructure according to AI: OpenAI, LangSmith and pgvector lead, Exa edges Tavily and Firecrawl
Key findings
- OpenAI was named in 100% of LLM API answers and first in 66%. Google was named in 100% and first in 4%. Anthropic was named in 92% and first in 28%. Anthropic's own models named OpenAI first more often than Anthropic.
- pgvector, Qdrant and Pinecone were named in 94% to 98% of vector database answers. Pinecone was first most often (38%), then pgvector (32%). Weaviate and Milvus: 90% share, 0% first.
- LangSmith was named in 100% of observability answers and first in 16%. Langfuse was named in 96% and first in 50%: the models list LangSmith and recommend Langfuse.
- Exa, Tavily and Firecrawl were named in 92% to 94% of search API answers. Tavily was first in 32%, Firecrawl 28%, Exa 8%. Firecrawl's site was the most cited source in the category at 233 citations.
- Playwright leads browser automation at 96% and 42% first. Browser Use (80%, 30% first) was the strongest agent-native tool; Browserbase was named in 82% and first in 2%.
The AI infrastructure categories are where the models rank their own makers, and they do not flatter them. Anthropic’s three models named OpenAI first more often than Anthropic. OpenAI’s models put Google’s Veo above Sora in video. Google’s models named OpenAI first in LLM APIs. The infrastructure layer is also where the leader on mentions and the leader on recommendation split most often.
LLM APIs
OpenAI 100% share, 66% first. Google 100% and 4%. Anthropic 92% and 28%. DeepSeek 66% and 2%. OpenRouter 54%. Amazon Bedrock 52%. All ten models agreed on OpenAI. Google is in every answer as the price-performance option and is the pick in two. Anthropic is the pick in 14 of 50, and its own three models split their firsts between OpenAI and Anthropic.
Vector databases
pgvector 98% and 32%. Qdrant 98% and 24%. Pinecone 94% and 38%. Weaviate 90% and 0%. Milvus 90% and 0%. Chroma 48% and 4%. The answer share is flat across five products; named-first separates them. Pinecone is the most frequent pick, pgvector the most frequent mention. Weaviate and Milvus are on 45 of 50 lists and open none.
LLM observability and evals
LangSmith 100% and 16%. Langfuse 96% and 50%. Arize 94% and 2%. Braintrust 92% and 14%. Confident AI 68% and 12%. Helicone 56% and 0%. The clearest inversion in the Index: LangSmith is named in every answer and recommended in eight; Langfuse is recommended in 25. Open source and self-hosting are the reason given. Confident AI’s own site was the top cited source in the category at 155 citations. PromptLayer, Humanloop, Laminar and Sentry were never named.
Inference hosting
Hugging Face 96% and 34%. Together 90% and 38%. Fireworks 86% and 8%. Baseten 72% and 0%. Modal 68% and 4%. RunPod 68% and 0%. Together is picked slightly more often than Hugging Face despite fewer mentions. GPT-5.6 Sol alone most-named Modal. SambaNova was never named.
Web search and scraping APIs for AI
Exa 94% and 8%. Tavily 92% and 32%. Firecrawl 92% and 28%. Brave Search 84% and 8%. Serper 62%. Bing 60% and 14%. Exa leads on mentions and is picked least of the top three. Claude Opus 5 alone most-named Firecrawl. Firecrawl’s own domain drew 233 citations, the most of any vendor site in the Index; it did not translate into the most firsts.
Browser automation for agents
Playwright 96% and 42%. Stagehand 82% and 14%. Browserbase 82% and 2%. Browser Use 80% and 30%. Skyvern 54% and 0%. Selenium 48%. Playwright, a Microsoft open-source library, leads a category of venture-backed agent tools. Browser Use is the pick in 15 of 50. Browserbase is on 41 lists and opens one. Claude Fable 5 most-named Stagehand; Sonar Reasoning Pro most-named Browser Use. Apify and Notte were never named.
What the field shows
Infrastructure is the field where “named in most answers” and “recommended” diverge most. LangSmith and Langfuse, Exa and Tavily, Hugging Face and Together, pgvector and Pinecone: in each pair the first is mentioned more and the second is picked more. The models cite documentation and comparison pages in this field, not YouTube; five of these six categories had no YouTube citation in their top sources.
What this means for a vendor
An infrastructure vendor’s answer share is table stakes; the top five in each category are within a few points. The named-first number is the market. It moves on the specific reason the models give (open source, self-hostable, cheapest, fastest), and that reason lives in documentation and comparison pages, the sources these categories cite.
Limitations
Fifty answers per category. “Anthropic” as an API vendor and Anthropic as a model provider are the same company measured from both sides; the overlap is reported, not adjusted. One edition.