Anthropic filed a confidential S-1 with the SEC last week, asking public markets to put a number on the future of closed-model premium pricing. Nathan Lambert, writing at Interconnects, published his framework for why that number is harder to set than it looks.
Lambert’s central claim is structural: open and closed models are not on the same capability curve at different positions. They are on separate exponentials optimizing for separate ends. Closed labs such as Anthropic and OpenAI are compounding at the high-margin, out-of-distribution end of the market, where complex reasoning, frontier code generation, and multi-step agents command a price premium that users demonstrably pay. Open-model builders, led by Meta, Qwen, DeepSeek, and MiniMax, are compounding toward breadth, cost, and tunability across a far larger surface area of enterprise deployments.
The bull case for Anthropic’s IPO rests on the first curve staying durable. Lambert grants it real foundation: coding agents past the Opus 4.5 and Codex 5.2 capability thresholds have shown that professional users will pay a sustained premium for better intelligence when their output depends on it. The integration advantage of closed labs, combining model weights, serving infrastructure, and hardware optimization, compounds in ways that commodity open-inference stacks cannot easily replicate. Lambert draws the Apple-and-Microsoft analogy: selling a tightly integrated, hard-to-replicate product while running high-leverage subscriptions across the economy.
The bear case is Lambert’s second exponential. Once open-model post-training infrastructure matures enough to close the out-of-distribution gap, enterprises that are currently stuck on closed models due to degraded performance on their hardest tasks will have a credible exit. Lambert puts foundation models from the open-weight tier within roughly six months of closed-lab capability on standard benchmarks. When they stop chasing Anthropic and GPT on the Artificial Analysis index and start filling the niche of reliable, low-cost deployment, the bottom of the market shifts permanently upward with each new weight release.
The asymmetry is the key tension for anyone pricing Anthropic’s shares. A closed lab must charge premium rates to fund the next training run; an open release distributes capability at no marginal cost once weights are published. Each open release resets the floor of what enterprises will accept to pay for. The closed-lab oligopoly does not need to lose the top end to lose margin power. It only needs the middle of the market to find open models adequate.
Lambert is not arguing that Anthropic and OpenAI will fail. His five-to-ten year valuation range of $2 trillion to $10 trillion per company is not a pessimist’s forecast. His argument is that the total value capture of the open-model economy, distributed across fine-tuning vendors, inference infrastructure companies, and enterprise integrators, will collectively exceed what the closed labs accumulate. Concentrated oligopoly on one side, diffuse and larger market on the other.
The S-1 puts the IPO pricing question in sharp relief. Anthropic’s prospective investors are essentially taking a position on whether the out-of-distribution gap remains wide enough, and durable enough, to justify a premium multiple through the public market lifecycle. If Lambert’s catch-up timeline of roughly six months on capability is close to accurate, the pricing window for that gap is not a five-year assumption. It is a near-term product execution bet.
Teams evaluating open-model switching timelines for their own infrastructure should benchmark the specific out-of-distribution tasks where their closed-model dependency is hardest to replace; that gap, not benchmark leaderboards, is the real signal for when Lambert’s second exponential starts compressing Anthropic’s pricing power in your stack.
Interconnects by Nathan Lambert (interconnects.ai), 2026-06-02.
