OpenAI Shipped Everything. Your Team Still Needs Trust and Portability.

The signal in the noise. OpenAI Dev Day shipped a lot: Dots (the persistent assistant), 6.1 Soul (a fast, cheap model that rivals Opus 5.5), and a decisions API that directly competes with Jev. The panel of experienced builders agrees these are technically impressive, but the real story is how hard it is to turn frontier models into software people actually trust with their data.

Where products break down. Dots is the clearest example. It's the right idea—an always-on assistant that learns from your behavior—but it's buried inside the ChatGPT app, not a standalone product. The performance of the Codex desktop app has degraded with every update. Multiple panelists report that the systems they rely on feel directionless, as if the labs are shipping features without a coherent product strategy.

The trust bottleneck. Connecting a model to personal email has become the litmus test. Most panelists connect their email only to OpenAI or Anthropic products, and even then with write-explicit-permission guards. Prompt injection may be largely solved on frontier models, but the more dangerous risk is the model incorrectly classifying what should or should not be shared when responding to emails.

The model tier that matters. 6.1 Soul is the surprise standout. It delivers near-Opus 5.5 quality at a fraction of the cost, with minimal reasoning overhead. For classification and structured output tasks, Jev-class models reduce costs 10-20x while maintaining accuracy. The panel notes that most teams still overpay by using reasoning-heavy models on problems that don't need them.

The infrastructure gap that remains. Execution portability between local and cloud is still unsolved. Coding agents keep state in memory, making it impossible to move work mid-task. Pi Durable solved this with an append-only log approach, but the big labs have not committed to the API. The absence of a standard for agent state transfer means every team builds their own harness.

The organizational takeaway. Engineering leaders should watch where trust and portability land, not model benchmarks. The panel's collective behavior—running parallel assistants, MCP gateways, and permission layers—reveals a reality where even the most sophisticated users cannot rely on any single product. That is the market signal worth paying attention to.