The Benchmark IPO Era: Reserved Frontier Models as Marketing Asset

A pattern emerging in late-cycle frontier model releases: labs increasingly publish benchmark numbers using internal, unreleased 'reserved' models alongside their actually-shipping product. Claude Mythos appeared in Anthropic's benchmark slides for the Claude Opus 4.7 release despite not being available to customers. SIHO, the human competitor who narrowly beat OpenAI's AHC system at AtCoder, summarized the dynamic: 'the benchmark game has entered its IPO era.' Both Anthropic and OpenAI are reported to be targeting 2026 IPOs, which gives both labs strong incentive to present the strongest possible benchmark story to public-market investors, even when the headline numbers reflect a model the buyer cannot actually access. This is structurally different from the 'our model is best' framing that dominated 2023-2025. The shift creates two evaluation problems for buyers and researchers: 1. Published benchmark numbers may not reflect the deployed model's behavior. 2. The reserved-model framing turns model evaluation into a marketing surface rather than a reproducible measurement. A related unverified theory, surfaced in Dwarkesh Patel interviews, proposes that some 'reserved' frontier models are kept unreleased partly as a NATSCmaxing argument — leverage in U.S. policy debates about restricting GPU exports to China on the grounds that the most capable American models are dangerous enough to warrant export controls. This is plausible-but-unconfirmed positioning, not documented strategy. See Gemma 4 Benchmarks and Performance for a contrasting open-weights release pattern, and GPU (Graphics Processing Unit): From Rendering Pixels to Training AI for hardware context.