Interconnects maps the next phase of model competition

Nathan Lambert used Interconnects to write a May outlook on the next phase of AI. It is not one news item, but a map of themes: Gemini Flash 3.5, Mythos, open-closed balance, agent tools and power shifts around open-source.

Open models have not yet had their true agent moment

The piece argues that 2026 will not have pauses in AI impact. Lambert describes rising model capability, fast changes in work, real AI economics and real-world risks coming forward.

One main thesis is that open models have not yet had their true agent moment like Opus 4.5 in Claude Code. Lambert proposes a practical test: not benchmarks, but whether open-weight models become genuinely useful in agentic harnesses. He also says Google still lacks a clear substitute for Claude Code and Codex.

The debate shifts from rankings to real work

This is a useful frame for teams choosing a model strategy. An open model can be cheaper, more controllable and suitable for enterprise agents. A closed frontier model can still win on robustness, workflow and products that people actually use every day.

Lambert shifts the debate from rankings to work. Whether a model clears a benchmark is one thing. Whether it survives as a tool in a long coding or agent workflow is the harder question.

Lambert's outlook is an analytical hypothesis, not a dataset

This is commentary and outlook, not a dataset. Lambert is a strong curator, but some claims are forecasts about future model specialization and lab economics. Read them as analytical hypotheses, not settled facts.

The claim about open models lagging will also depend on the use case. For cheap automated tasks, an open stack may be good enough earlier than it is for top-end knowledge work.

The first open-weight model with Claude Code reliability will be the key signal

The most important signal will be the first open-weight model that gives developers a Claude Code or Codex level of reliability. Not in a tweet, but across a working week.

The second line is Google. If Gemini does not quickly get a strong coding and agent product, Google can have excellent models and still lose the most visible work workflow.

Lilith's verdict

Lambert's piece is less a prediction and more a checklist. Anyone waiting for one winning model is standing in front of a board where every arrow points in a different direction.