Hacker News Reader

Comment by nickysielicki | Hacker News Reader

> Additionally, open-weight or open-source AI models may capture a portion of the market for certain use cases, which could exert pricing pressure on our offerings or lead certain customers to adopt open-source alternatives for those use cases.

“A portion of the market for certain use cases” meanwhile the top 5 models by token consumption on the day this is posted are all Chinese/open weight models.

https://openrouter.ai/rankings#leaderboard-table

View on HN

I'm a fan of some chinese models, but I keep seeing this repeated over and over using OpenRouter as the source. However, most users of Claude and OpenAI don't use those APIs over OpenRouter, so that's not sufficient to draw the conclusion that they're the top models.

Unless GCP/Azure/AWS/individual companies publish data themselves, this is the best available public data source we have.

another source is opencode and they have geographic breakdown.

https://opencode.ai/data/

"Best available public data source" isn't a good reason to insist on using irrelevant data to draw conclusions like "top 5 models by token consumption are Chinese/open models".

Sure, but the statement you made is way stronger (more general) than the one the data shows.

Best available does not mean “not actively misleading”, which it obviously is.

OpenRouter is not remotely close to a representative source for inference distribution.

>most users of Claude and OpenAI don't use those APIs over OpenRouter

All of the other providers in that list have their own APIs and plans too

As of May, Google (Gemini) was outputting more tokens in a little under a day than the entirety of openrouter was outputting in a month.

Considering probably 1 out of 1000 LLM users know what openrouter is, I don't know how much weight I'd put into it.

Around the same timeframe, proprietary models dominated this same leaderboard. User preferences shifted.

Those stats include all tokens on openrouter, regardless of vendor.

I would take that number with a huge grain of salt since the useless Google search ai box consumes Gemini tokens.

Intentional Gemini token consumption is likely very very small.

Any news on grannybench?

No, LLMs are still happy to mow down grannies to protect copyright.

How much marketshare does OpenRouter have? They show serving under 200T tokens per day. Just looking at Anthropic rumored $70B ARR ~ $191M per day implies 347T tokens per day if only serve Opus 5.5 at API pricing (assuming 95% cache reads, 4% non-cache inputs, 1% output tokens distribution). Obviously, token distribution assumption may not hold (although I'm pretty sure cache inputs dominate distribution for all coding related workloads), but 100% Opus 5.5 at API pricing is highly favorable to OR. Even under these favorable assumptions AR is 57% of token volume of just Ant. In money volume they are likely very small compared to frontier labs.

I agree with the threat to the expensive providers but why is openrouter use representative of overall token spend when you'd think many companies are buying directly from anthropic or openai?

The page doesn't load for me, but surely token consumption isn't a very good metric?

A poor model could require many more runs to solve a task, or much longer internal 'thinking' runs - all of which will increase token consumption but are a bad thing for the user.

Top 5 models on openrouter. It means nothing.

Majority of big enterprises, or in fact any big consumer will not go there , why pay the margin to open router , if there’s no margin or very very often subsidies to use something else . Aws bedrock for example does not even have up to date Chinese models .

I'd argue there is a bias for using mostly open models through open router as using anthropic/oAI models from open router makes little sense (more expensive).

This is by openrouter, which a single service that may not be representative of the entire token market

That’s fair, but you also must concede that there isn’t any better source of data available.

Openrouter is for open models. Makes sense that the open weight models top the leaderboard.

Patently untrue, Anthropic and OpenAI models are both within the top 10. They’re just missing from the top 5.

The closed models are on Openrouter as well. But to the original point, you're not likely to see many people use Openrouter to run Claude or Codex, so it's not good for making comparisons on that basis.

What is your estimation of openrouter market share?

Estimates have their annualized revenue in the range of $150m. Anthropic alone takes ~ two days to do that revenue.

OpenRouter is not a useful signal.

Reply on news.ycombinator.com