Frontier AI Labs Score High but Fall Short on Software Readiness

Summary

Opsera’s Frontier AI Software Readiness Benchmark applies its eight-dimension ForgeScore to 83 public repositories maintained by 29 organizations building frontier AI models. The repositories were assessed as they appeared in August 2026; private repositories, internal systems, and overall engineering performance were outside the scope. ForgeScore rates code from 0 to 100, with 75 or higher considered strong, 40 to 74 moderate, and below 40 weak. The average score was 74.6, just below the strong threshold. OpenAI recorded the highest organization-level score at 78.6, while scores across the 29 organizations ranged from 67.0 to 78.6 and no organization reached the 80s. The results suggest that access to advanced models, engineers, and compute does not by itself make software easier to understand, govern, secure, or sustain. Trust Boundaries was the weakest of the eight dimensions, averaging 71.1 against the 74.6 composite average; it covers authentication, authorization, secrets, exposed interfaces, service edges, and data access. The assessment cites repositories containing API keys in Meta Android client code and DeepSeek inference endpoints without built-in throttling or rate limiting. The article presents the findings as a shared industry constraint rather than a judgment on any individual lab: AI can increase the speed and volume of code changes, but software readiness still depends on context, specifications, safeguards, and governance.