More than half of the tokens flowing through OpenRouter now route to open-weight models, and the five busiest entries on the routing platform’s live leaderboard are open, according to The State of Open Source AI, a report on the open-model ecosystem published July 14. Anthropic’s Claude models are the next entrants built by a US lab. The frontier still belongs to closed systems. Production volume does not.
That split did not happen overnight. Per the report’s citation of an OpenRouter study covering 100 trillion tokens between November 2024 and November 2025, open weights held a negligible slice of routed traffic in late 2024, climbed to roughly a third by the end of that year, and crossed 50 percent by mid-2026. The gain concentrates in coding and agentic workloads, where developers route to whatever is cheapest per call. By request count rather than token count, closed US providers still lead, a distinction the report is explicit about.
Cost explains most of the shift. The report tracks a 50-fold drop in GPT-4-class inference pricing over 36 months, from $20 to $0.40 per million tokens. On a separate OpenRouter sample spanning May to September 2025, closed models held about 80 percent of usage but 96 percent of revenue, because at roughly 90 percent capability parity, closed inference still ran about six times more expensive per call. A Linux Foundation-cited estimate (the Nagle-Yue study) puts the resulting unrealized annual savings from switching to open weights at approximately $24.8 billion.
The capability story is narrower than the cost story. On Chatbot Arena, the open-versus-closed gap fell from 8.04 percent in January 2024 to 0.5 percent by August 2024, and DeepSeek-R1 briefly matched the leading US model in February 2025. By March 2026 the gap had widened back to 3.3 percent as closed reasoning models pulled ahead. The report frames that 3.3 percent as an average masking a jagged line: open sits at or near parity on coding, instruction-following and general knowledge, while the gap concentrates in reasoning, long-context retrieval and agentic planning.
None of this means open weights are simply cheaper and worse. It means the two camps have stopped competing on the same axis. Closed labs are defending a shrinking category, tasks that genuinely require frontier reasoning or multimodal judgment, while open weights have absorbed the much larger category of tasks that just need a fast, cheap, good-enough model wired into a workflow. The report’s own data on where teams stall, only 51 percent of open-model deployments reach production versus 63 percent for closed, according to its cited Mozilla/SlashData developer survey, suggests the constraint on open adoption is tooling and operational maturity, not model quality.
For a team deciding build-versus-buy today, the practical rule this report supports is: default to open weights for anything that resembles high-volume production traffic (coding assistants, agent loops, customer-facing automation), and reserve closed frontier APIs for the narrower slice of tasks where reasoning depth or multimodal accuracy is the actual bottleneck. Budget the extra engineering effort for the former; budget the per-token premium for the latter.
Figures and analysis are drawn from The State of Open Source AI, V1.0, published July 14, 2026 at stateofopensource.ai.
