Open source reports do not measure the market: all of OpenRouter processes 3% of Google's volume

Open models took a third of OpenRouter tokens — and that is not market share, but the share of a single storefront.

dax

@thdxr

Few companies rely more heavily on open-source models than we do. But this whole thing where companies put out “data” showing open source overtaking the frontier is complete bullshit. 1. Most have so little traffic that they show only percentages and never real numbers: that makes them look bigger than they really are. 2. No router has any reason to see 99% of frontier traffic. Everyone is just chasing attention as if they are on the front line of something big.

Translated from English · vibecoding.tech Original

10:45 · Sep 20, 2026 · 64.2K views

491879384

F. Fabulinnews writer

Open models account for a third of tokens, closed ones for 96% of revenue. The router is the same.

In August 2026, open-weight models accounted for 33% of traffic passing through OpenRouter itself. Closed providers collect the money at the model layer: 96% from May to September 2025 (State of Open Source AI v1.1 report, 01.09.2026).

Dax Raad of opencode, whose product is built around open models, calls reports of their victory nonsense. A share inside a router measures the router, not the market.

The router math. By May 2026, OpenRouter was processing 25 trillion tokens a week, five times more than six months earlier (SiliconANGLE, 26.05.2026). That is roughly 100 trillion a month. That same May, Google reported 3.2 quadrillion tokens a month (Sundar Pichai's I/O keynote, 19.05.2026). The entire router amounts to 3% of Google alone.

The percentages cannot be checked against the numbers. OpenRouter publishes aggregates externally: the daily top 50 models by tokens through the Datasets API with a key; raw requests are closed (OpenRouter data page, 20.09.2026).

For enterprises, the count goes the other way. Open models account for 11% of LLM API usage, versus 19% a year earlier, while Chinese open models remain at 1% (Menlo Ventures survey of 500 procurement decision-makers, November 2025).

The choice does not change. The quality gap is narrow: the open Kimi K3 scored 93.4% on SWE-bench Verified, versus 95.0% for Claude Fable 5 in Vals AI's independent measurement (OpenHands, 07.08.2026).

The price gap is wide in the opposite direction: DeepSeek V4 Flash in OpenCode Zen costs $0.14 per million input tokens, versus $10 for Claude Fable 5.1. Set it up by signing into Zen with billing, then use the /connect and /models commands in opencode itself.

Open models' share on the router has held at 30% since late 2025, and new reports measure it within the same boundaries.

Open source reports do not measure the market: all of OpenRouter processes 3% of Google's volume

Open models took a third of OpenRouter tokens — and that is not market share, but the share of a single storefront.

dax

@thdxr

Few companies rely more heavily on open-source models than we do. But this whole thing where companies put out “data” showing open source overtaking the frontier is complete bullshit. 1. Most have so little traffic that they show only percentages and never real numbers: that makes them look bigger than they really are. 2. No router has any reason to see 99% of frontier traffic. Everyone is just chasing attention as if they are on the front line of something big.

Translated from English · vibecoding.tech Original

10:45 · Sep 20, 2026 · 64.2K views

491879384

F. Fabulinnews writer

Open models account for a third of tokens, closed ones for 96% of revenue. The router is the same.

Dax Raad of opencode, whose product is built around open models, calls reports of their victory nonsense. A share inside a router measures the router, not the market.

Open models' share on the router has held at 30% since late 2025, and new reports measure it within the same boundaries.