Open Source AI Has Already Won the Performance Battle — Now It's About Distribution
TexTak forecasts a 69% chance that open-source models will match closed frontier performance, up from 67%. April's release cycle has effectively settled this question — GLM-5.1 now tops the SWE-Bench Pro leaderboard under an MIT license, beating GPT-5.4 and Claude Opus. The real question isn't whether open source will catch up anymore, but whether the closed labs can maintain any meaningful moats.
Our 69% reflects three converging trends: dramatic cost reductions in training, open research accelerating technique diffusion, and Meta's strategic commitment to open-source leadership. What changed our thinking this month wasn't just GLM-5.1's benchmark performance — it was the sustained velocity of open releases. Seven major models launched in the first twelve days of April alone, with multiple teams now achieving frontier-class results.
The counterargument centers on unreleased capabilities at frontier labs. Anthropic's Claude Mythos Preview, which discovered thousands of zero-day vulnerabilities including a 27-year-old OpenBSD flaw, suggests significant capability overhang that benchmarks don't capture. This is the gap in our model — public benchmarks may systematically underestimate the performance delta between open and closed systems because frontier labs are holding back their most capable releases.
However, the distribution advantage increasingly favors open models. LangChain reports that models like GLM-5 now match closed systems on core agent tasks at a fraction of the cost and latency. When performance parity meets 10x cost advantages, adoption momentum becomes overwhelming. The question shifts from "can open source catch up" to "how long can closed labs justify their premium."
What would move us above 75%? Evidence that frontier labs are running out of proprietary techniques to withhold, or that open training budgets are scaling faster than we've modeled. What would drop us below 60%? Proof that benchmark gaming is inflating open-source performance, or that closed labs have achieved genuine architectural breakthroughs that can't be reverse-engineered from published research.