Open and closed models are on different exponentials

Nathan Lambert wrote this short piece for interconnects in June 2026. His claim is that the balance of power between open and closed models will be settled by economics, specifically by whether users keep paying large margins for the best closed models. Coding agents gave the first answer in early 2026: at least one huge market will pay a substantial premium for better intelligence. The open-source-ai-reading-list files the post as the explanation of where adoption differs between the two.

The closed exponential

The habit change set in with coding agents after the Opus 4.5 and Codex 5.2 thresholds. Lambert says people switched because their net output is higher with an agent as an implementation aid, and anyone who relies on agents for work will pay for the best rather than settle for good enough. He would pay $2,000 a month for today's tools. He also grants that many companies push agents on employees who get little from them in their current form, which keeps the buildout, or the bubble, going.

The best closed labs, currently Anthropic and OpenAI with Google likely to catch up, will always make the most efficient models for a given level of intelligence. Weights, harness, tools and serving infrastructure together have large returns to integration, and open models are designed to work across many serving setups instead. If benchmark scores saturate, labs will optimize utility per second or per watt; there have been no walls in progress. Lambert agrees with the labs' focus on raw intelligence because it opens new markets, while optimizing at a fixed intelligence level locks existing markets in and lowers margins.

His analogy is the iPhone. You could buy an Android and put up with paper cuts to save money, but why would you, and at work the returns to performance are higher still. He expects the frontier labs to look like a mix of Apple, selling integrated technology that is very hard to replicate, and Microsoft, selling high-leverage subscriptions across the economy. In 5-10 years he expects OpenAI and Anthropic each to be valued at $2-10 trillion, with the true frontier an oligopoly resembling today's cloud market.

The flip side is a slow decay of the labs' API businesses. They will protect their best models by rolling them out to APIs later, to conserve token supply, avoid distillation, and stay with high-margin uses. Lambert expects this to be clearly visible over 5-10 years; in the near term, supply-limited compute and heavily subsidized tokens set prices, margins and demand.

The open exponential

The collective value around open models will be far bigger, Lambert argues, and will dramatically exceed the combined value of OpenAI and Anthropic, but revenue and margins will be spread across a wide stack of companies. Many businesses want to switch to open models, and today those models are not good enough on out-of-distribution tasks. Eventually open builders will stop chasing Claude and GPT on the Artificial Analysis index and fill that gap, either because they can no longer fund rising R&D costs or because some products can only exist at open-model prices.

Open models are not integrated, so several companies have to coordinate to serve them, every layer has alternatives, and prices fall toward commodity levels. Low, predictable prices are where enterprises start building in-house agents and tools for niche tasks. The usual deployment is to find a model that clears a performance bar on a task and never replace it, because setup is costly. Fine-tuning services such as Tinker, Fireworks and Prime Intellect make that market larger. He predicts a steady rise in the share of open-model inference on the Google, Amazon and Microsoft clouds and on Together, Fireworks and OpenRouter, relative to OpenAI and Anthropic.

The two economies run on different exponentials. Progress stays fast across the board, and Lambert calls claims that recursive self-improvement will give closed labs an unassailable lead overblown. New products such as background agents can run on either kind of model. Closed models found product-market fit at the top of knowledge work first; the open economy will take far longer, because it tracks the spread of AI through the whole economy.

How it fits with the other pieces

The post turns the argument of what-comes-next-with-open-models and the gap data in open-models-in-perpetual-catch-up into a business forecast. Christian Catalini's some-simple-economics-of-open-versus-closed-ai, two months later, reaches the same split, premium tokens for a few buyers and commodity tokens for the bulk, from the economics of complementary assets. rl-finetune-beats-frontier is a concrete case of the "good enough on one task, never replaced" deployment Lambert describes.

The iPhone analogy puts Lambert at odds with Bill Gurley's from-open-source-software-to-open-source-strategy, published a month earlier. Gurley also casts OpenAI as Apple in 2008, but as the incumbent a coalition of followers will commoditize the way Android did. Lambert agrees the open side ends up with more total value, which is Android's outcome, and still expects the premium tier to keep its pricing power and grow into trillion-dollar companies. The disagreement is over how much of the market the Apple position keeps.

The forecast leans on one market. Coding agents are the only evidence offered that users will pay large premiums for the best model, and the post does not ask whether that willingness extends to other kinds of knowledge work.