My bets on open models, mid-2026
Interconnects Nathan Lambert
An AI researcher lays out ten predictions on open vs closed AI models through mid-2026. Bottom line: it's not about ability anymore, it's about money and who keeps funding the models.
Based on reporting by Interconnects, Nathan Lambert — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nathan Lambert, the researcher behind Interconnects, just published his clearest attempt yet at untangling one of AI's messiest debates: will open-weight models ever really catch closed ones like GPT and Claude. His answer is a firm no, but not for the reason most people assume.
The surprising part of his argument is that open models have actually held their own on benchmarks through late 2025, despite closed labs having far more compute and research firepower. Chinese labs in particular have gotten very good at chasing frontier scores, partly through distillation techniques that let them fast-follow whatever the big U.S. labs release. Lambert says that's not going away even if regulators try to choke off distillation access, because it's now baked into how these labs raise money and win adoption.
Where closed models still win, he argues, is in the messier, harder-to-benchmark stuff: robustness, reliability, and handling the weird edge cases that come from millions of individual users hammering a product like Claude or ChatGPT every day. He thinks that gap will widen specifically in agentic coding tools, Claude Code and Codex being his examples, because closed labs can tune models using live feedback from real users in ways open labs simply can't replicate.
The funding angle is really the spine of the piece. Lambert expects Chinese open-weight labs to hit financial trouble as early as later this year, with the effects showing up in slower capability gains three to nine months after that. He also predicts the U.S. will claw back ground in open-model adoption starting in early 2027, pointing to Google's Gemma line, Nvidia's Nemotron, and startups like Arcee AI as evidence the pendulum is already swinging.
He throws in a wildcard too: personal AI agents and local tools, which he calls a kind of dark matter for the whole ecosystem, mostly invisible in current metrics but potentially massive in shaping who wins. And he's blunt that attempts to ban powerful open models outright will fail, since training near-frontier systems is cheap enough that someone, somewhere, will just release them anyway.
My take — AI-written commentary, not fact-checked reporting
I'll take the unglamorous economics argument over the geopolitical panic any day. Everyone wants a simple story where China either wins or loses the open-model race, but Lambert's right that funding runways and who actually deploys these things matter more than benchmark bragging rights. The real tell will be whether anyone builds a funding model for open weights that doesn't depend on one company's goodwill or one government's subsidy.
Read more about this at: Interconnects