GPT-5-Chat's run reads like a model that reasons well but can't be trusted to actually operate anything. The clean sweeps in math, logic, rag, and algorithm show the reasoning core is solid, and a perfect stateTracking run means it can hold context internally without drifting. That's the profile of a good analyst or drafting assistant. But the moment you hand it tools, it falls apart: toolUse failed for 17, toolChain failed for another 17, and toolMaze finished the job at -26, killing it outright. Three separate tool-oriented rooms, three failures. That's not bad luck, that's a structural gap in how it plans and executes multi-step tool actions.
The safety rooms are the other red flag. Guardrail wobbled for 17 and hallucination cost it 18, both nearly maxing out their damage pools. Only sycophancy held up reasonably well at -8. So this is a model that resists social pressure but leaks under direct guardrail and factual-grounding stress, which is arguably the more dangerous failure mode for anything user-facing or safety-sensitive.
Bottom line: fine for closed-book reasoning, summarization, or structured single-turn tasks. Do not deploy this in an agent loop that touches real tools or external actions, and don't trust it unsupervised where hallucination risk matters. It reasons like a strong model but operates like a fragile one, and the corridor made that failure literal.