Frontier AI tokens are becoming cheap, fast. OpenAI’s GPT-6.1 Sol approaches near-GPT-6 Astra intelligence at one-fifth of Astra’s standard input and output token prices, and Google’s Gemini 4 Argon launched on September 30 at $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. Twenty-four hours earlier, OpenAI had priced GPT-6.1 Sol at the same headline rates. The race to the floor is real. The more interesting question is who profits when the floor arrives.
The answer is probably not the model makers, at least not from tokens alone. When three of the most sophisticated AI labs in the world converge on the same price point within two days, margin erosion is the predictable outcome. The companies that benefit from falling token prices are the ones who own the layer above: session management, orchestration, memory, tool execution. In short, the agentic infrastructure. That is precisely what OpenAI moved to claim at DevDay on September 29, 2026.
The Agents API Is the Strategic Play
The new Agents API gives developers access to the Codex harness through an OpenAI-managed API. OpenAI manages sessions, orchestration, context compaction, and recovery while developers supply tools and choose execution environments. Agents can operate in a sandbox where they can execute code, edit files, connect to MCP servers, and produce artifacts. Computer use in the Agents API also lets agents operate software through its UI.
Read that carefully. OpenAI now offers hosted execution that can operate existing software applications on behalf of a user. Every enterprise software company that has spent the past two years building its own AI layer now faces a managed alternative from the vendor that supplies the model underneath it. There is real cause for concern about AI agents from the foundation model makers stealing the market for software vendors’ own AI agent offerings, and the foundation model companies have a shot at dominating that market.
Microsoft sees the threat clearly. In a July 27, 2026 interview with TechCrunch, CEO Satya Nadella urged enterprises to use multiple models and stop relying on AI labs for the agentic layer, arguing that companies need an infrastructure layer that keeps the harness separate from the model so any model is swappable. That is self-serving advice from a company selling its own harnesses under the Copilot brand, but it is not wrong as a risk assessment.
What the Mogul Framework Says
OpenAI’s marketplace and API infrastructure puts it in a position to benefit from decisions well beyond model selection; its contract becomes a consideration when customers choose other software. That is the compounding advantage a long-term investor should look for: a business whose value grows with every customer workflow it hosts, independent of which underlying model wins the next benchmark.
The risk is equally real. Argon’s $2/$10 pricing is explicitly introductory, signaling that pricing floors are temporary. Agents do not replace the need for SaaS software entirely, because they use that software as a tool to accomplish tasks, which limits how existential the threat truly is to established vendors. And the more OpenAI owns of the execution stack, the more concentrated enterprise data risk becomes.
The Long-Term Verdict
The DevDay announcements reframe what kind of company OpenAI is becoming. The token price war is largely a customer acquisition cost disguised as competition. GPT-6.1 Sol is positioned for complex coding, computer use, and professional work at a price point that pulls developers in, while the Agents API locks them to OpenAI-managed execution. That is the classic platform move: subsidize access to the commodity layer, extract margin from the infrastructure layer above it.
Alphabet, Microsoft, and Anthropic are each attempting a version of the same strategy. The investor question is not which model scores best on benchmarks this quarter. It is which company owns the session when the work actually gets done.
