GPT-5 will not have reliable models of what it talks about that external programmers can access to feed downstream processes.

Judge GPT-5 as released. True: GPT-5 offers external programmers no explicit, inspectable internal model of the entities or situations it discusses (such as a queryable world-state or knowledge structure) that downstream software can rely on. Programmers can only get model-generated outputs, and independent tests show these remain prone to errors that make them unreliable for downstream use. Formatting features such as JSON or structured output schemas do not count as accessible models of things. Mostly true: the above holds, but OpenAI offers some limited, partial mechanism exposing internal representations, and that mechanism is not reliable. Mixed: there is an accessible mechanism that is reliable in some domains but not others. Mostly false: there is an accessible and largely reliable explicit model interface, with minor failures. False: GPT-5 provides external programmers with reliable, accessible explicit models of what it discusses, and independent evaluation shows downstream processes can depend on them.

Set before the research began.

1. E1 Introducing GPT-5 for developers OpenAI · 7 August 2025

   “We’ve also added a new tool type—custom tools—to let GPT‑5 call tools with plaintext instead of JSON. Custom tools support constraining by developer-supplied context-free grammars.”

2. E2 Introducing GPT-5 for developers OpenAI · 7 August 2025

   “On prompts from LongFact and FactScore benchmarks, GPT‑5 makes ~80% fewer factual errors than o3.”

3. E3 Introducing GPT-5 for developers OpenAI · 7 August 2025

   “As with all language models, we recommend you verify GPT‑5’s work when the stakes are high.”

4. E4 Introducing GPT-5 for developers OpenAI · 7 August 2025

   “They also support parallel tool calling, built-in tools (web search, file search, image generation, and more), core API features (streaming, Structured Outputs, and more), and cost-saving features such as prompt caching and Batch API.”

All the evidence comes from OpenAI's developer page. It describes custom tools, grammar constraints and Structured Outputs [E1][E4], which the criteria count as formatting features. It also reports fewer factual errors than o3 [E2] and advises verifying high-stakes work [E3]. No evidence item shows whether GPT-5 offers an accessible internal model, and no independent tests of downstream reliability were provided, so the criteria for a verdict are not met.