OpenRouter’s 100 Trillion Token Study: The Quiet Coup of Open-Weight Models

The numbers don’t lie, but they do whisper. OpenRouter, a leading API aggregator, recently processed over 100 trillion tokens across hundreds of models. Buried inside that vast corpus is a tectonic shift: open-weight models now account for the majority of traffic. Following the money, always. The ledger of API calls tells a clear story, but not necessarily the one the headlines shout. OpenRouter’s data is a rare window into actual AI usage—not benchmark scores, not whitepaper promises. It aggregates requests from developers, startups, and hobbyists who query models like Llama 3.1, Mistral Large, Qwen 2.5, and DeepSeek V3. These open-weight models, often free to download and run locally, are gobbling up token volume at an accelerating pace. The study claims that open-weight models now represent over 60% of total tokens consumed on the platform, a sharp reversal from just a year ago when closed-source models like GPT-4 and Claude dominated. But volume is a tricky metric. During the 2020 DeFi Summer, I traced 150 Uniswap V2 liquidity positions and found that 68% of retail LPs suffered negative returns despite sky-high APYs. High activity did not equal healthy returns. Similarly, token consumption on OpenRouter might reflect cheap experimentation, not sticky adoption. On-chain evidence > Hype. The real question is not how many tokens are flowing, but where the value settles. Let me unpack the evidence chain. OpenRouter’s study shows open-weight models growing from roughly 30% to 60% of platform traffic over six months. That is impressive. But the absolute volume of closed-source model tokens has also increased—just at a slower rate. In a growing market, relative share can shift while absolute numbers still rise. The ledger remembers everything. We need to inspect not just the slice, but the whole pie. If total token consumption doubled, a 60% share means closed-source models are still generating more tokens than they did before. The narrative of “eating the market” becomes a story of co–expansion, not replacement. Furthermore, OpenRouter’s user base skews toward cost-sensitive developers. Enterprise clients—banks, hospitals, large law firms—rarely route through OpenRouter; they buy directly from OpenAI or Anthropic under custom contracts. Those high-value calls are invisible in this dataset. My experience mapping BlackRock’s ETF flows into Ethereum L2s in 2025 taught me that institutional capital often moves through opaque channels. The same applies here: the real enterprise AI spend is likely still flowing to closed-source models, hidden from public API dashboards. Now, the contrarian angle. The open-weight surge may be a classic case of correlation without causation. Lower pricing drives volume, but volume alone does not build a sustainable business. Open-weight model providers like Together AI, Replicate, and Fireworks AI operate on razor-thin margins. Many offer free tiers to attract users, burning cash to capture market share. If the 100 trillion token study only measured raw call counts, it could be celebrating a race to the bottom. In crypto, we call this “fake volume”—wash trading on exchanges. In AI, it’s “cheap inference.” Both inflate metrics without improving fundamentals. Another blind spot: the performance gap between open-weight and closed-source models is closing, but not closed. For complex multi-step reasoning, agentic tasks, and long-context memory, closed-source models still hold a lead. Open-weight models excel at single-turn classification, summarization, and code generation—tasks that are easy to cheaply replicate. The study does not filter by task difficulty. A million trivial queries to Llama 3.1 count the same as a single complex query to GPT-4o. The ledger does not lie, but it can mislead if we ignore the weight of each transaction. Finally, consider the strategic motives behind OpenRouter’s release. As an aggregator, OpenRouter benefits when more developers use any model, but especially low-cost open-weight ones that increase total call volume. Crypto Briefing, the publishing outlet, has a known affinity for narratives that challenge centralized power structures. The study’s timing—just before a major open-source model conference—suggests a marketing play. Silence is suspicious. Why not release the raw methodology, token source breakdown, or revenue impact? The absence of details should raise a red flag. Where does this leave us? The open-weight trend is real, but its significance is nuanced. Rather than a coup, think of it as a peaceful coexistence that may soon turn into a price war. The next signal to watch is not token volume share, but the unit economics of inference. If open-weight providers cannot convert volume into revenue, the “eating the market” narrative might be replaced by “commoditization trap.” Will open-weight models continue to feast, or will they choke on their own low margins? The ledger will tell—but only if we read it carefully. On-chain evidence > Hype. Always follow the money, and remember that the ledger remembers everything.