Where AI goes from here, in Nathan Lambert's mid-2026 read

Nathan Lambert posted his mid-2026 read of where the AI field is heading, and the picture he sketches in Some ideas for what comes next is more sober than most year-ahead pieces. His central claim is that open-weight models still have not matched the agentic step-change Claude Opus 4.5 created with Claude Code back in December 2025. Lambert puts the gap at five to six months, and warns it could stretch to twelve before an open model catches up on real coding-agent work.

The same gap shows up at Google. Even with Gemini 3.5 Flash on the bench, Lambert argues there is still no real competitor to Claude Code or Codex for serious knowledge work, which pushes open models toward enterprise automation rather than developer tools. He does not expect an open-weights Mythos this year either, citing resource limits at Chinese labs (Kimi, Z.ai, DeepSeek, Qwen) and a more conservative safety stance from their corporate parents.

What he sees more positively is American open work picking up: Nvidia's Nemotron and Google's Gemma 4 are stabilising the US open ecosystem, and Gemma 4 already matches or beats equivalently sized Qwen 3.5/3.6 models. He also flags that established power structures are noticing. The Pope's 40,000-word encyclical, China's new travel restrictions on AI researchers, and the US designating Anthropic a supply-chain risk all happened in the last few weeks.

Why it matters

If you are choosing between an open model and a frontier API for an agent project right now, Lambert's read says the open side is still too far behind for serious coding work, but plausible for narrower enterprise automation. That is a useful calibration before you commit a quarter to the wrong stack.