Open-Source LLMs Overtook Proprietary in 2026

Three months ago, proprietary models ran the show on OpenRouter. In mid-March 2026, the four big closed labs (Anthropic, Google, OpenAI, and xAI) took roughly 59% of every token routed through the marketplace, against about 41% for open-weight labs. By June 17 that had reversed almost exactly: open-weight labs took 60%, the closed camp 40%. On a seven-day average the lines crossed in early June, and a few volatile single days had already flipped in late May.

The data comes from Dirac’s Token Market Share by Lab dashboard, which compiles OpenRouter’s published daily token usage across every provider it routes to and groups it by the lab that built the model. The window below is 91 days, March 19 to June 17, 2026. The latest day is partial and dropped, consistent with Dirac’s stated method.

How open source pulled ahead

The flip is not a story of proprietary usage collapsing. It is a story of open-weight usage growing faster off a marketplace that was itself exploding. Total tokens routed through OpenRouter nearly tripled over the window, from about 2.2 trillion per day in mid-March to roughly 6.0 trillion by mid-June. Proprietary volume kept rising in absolute terms the whole time. Open-weight volume just rose much faster, so it won the ratio.

The mechanism is price. A routing marketplace like OpenRouter selects hard for cost-sensitive demand: developers and apps that send work to whichever model clears the job cheapest, switching providers through one API. That is the exact audience open-weight models court. The Chinese open-weight cluster (DeepSeek, MiniMax, Moonshot’s Kimi, Z.AI’s GLM, and Xiaomi’s MiMo) reached a tier of capability through 2026 that handles most production traffic at a fraction of the per-token cost of frontier closed models, a shift this site covered in why open-weight LLMs got good enough to matter in 2026. When a good-enough model costs an order of magnitude less per token, a price-sensitive marketplace routes volume to it. The same cost gap drives the spreads in what the leading AI coding plans actually cost.

The labs behind the shift

One lab carried most of the swing. DeepSeek went from 8.6% of OpenRouter tokens in March to 22.4% by mid-June, making it the single most-used lab on the platform, ahead of every closed competitor. Xiaomi was the other big mover on the open-weight side, climbing from roughly 3% to around 11% with sharp spikes along the way. MiniMax and Z.AI held high-single to low-double-digit shares; Moonshot and Mistral stayed smaller.

On the proprietary side, the decline was uneven. Google lost the most ground, sliding from about 20% to 12%, and OpenAI fell from 14% to around 8%. xAI nearly disappeared from OpenRouter routing after mid-May, dropping from roughly 3% to a fraction of a percent. The exception is Anthropic, which held near 20% for the whole window and remained the largest proprietary lab and the second-largest overall behind DeepSeek. Whatever pressure open-weight pricing put on the marketplace, Anthropic’s traffic stuck.

What the OpenRouter signal does and does not mean

Read the headline precisely: open-weight models overtook proprietary models among the tokens that flow through OpenRouter. That is a real, large slice of developer-facing LLM traffic, and OpenRouter’s own State of AI study put open-weight usage near a third of the platform by late 2025, so a move past half in 2026 fits the trajectory. But it is not a claim about all AI usage, and three caveats keep it honest.

First, this is a routing marketplace, not the whole market. The largest closed labs sell most of their volume through their own APIs and consumer apps (ChatGPT, Claude, Gemini), none of which passes through OpenRouter. The platform structurally undercounts proprietary’s true share of global LLM usage.

Second, this is token volume, not revenue. Open-weight models are far cheaper per token, so a 60% token share converts to a much smaller share of dollars. The camp that lost the token race is still collecting most of the money.

Third, “open source” here mostly means “open-weight,” and that camp is overwhelmingly Chinese: DeepSeek, MiniMax, Moonshot, Xiaomi, and Z.AI account for nearly all of it, with France’s Mistral a rounding error. The open-versus-closed story and the China-versus-US story are, on OpenRouter in 2026, almost the same story.

Sources

Dirac Delta Labs (2026). Token Market Share by Lab. Dirac, compiled from OpenRouter aggregate daily token usage. https://dirac.run/labs-market-share

OpenRouter (2026). LLM Rankings. OpenRouter. https://openrouter.ai/rankings

OpenRouter (2026). State of AI: An Empirical 100 Trillion Token Study. OpenRouter. https://openrouter.ai/state-of-ai