Post

MCP analytics is the new web analytics. We made ours public, and built a leaderboard of the most popular models, harnesses, tools, and more we see at PostHog. Here's what it shows:

Everyone loves Opus 5.5. Since its launch on September 22nd, Opus 5.5 has been by far the most popular model, making up nearly the majority of usage.

Opus 5.5 stands out because OpenAI models are split between GPT 6, 6.1 and 6.1-astra, but combined OpenAI (32%) is still behind Anthropic (65%). Interestingly, when it comes to Claude Code vs OpenAI (Codex), that's much tighter (32% to 19%).

Agents make a lot more calls than humans do, and they make calls in ways we don't expect. That leads to a lot of errors. A 3.2% error rate is terrible for normal requests, but given the special circumstances of agents, we don't think it's too bad (but we're obviously trying to make it better).

You might ask: How are we getting this data? Beyond capturing tool calls with MCP Analytics, the agents tell us in 73% of cases, metadata helps in 5%.

Also, in 93% of cases, agents are telling us why. This lets us make better tools for agents. It's like a mini user interview.