GPT-5.4 aces recall, flunks restraint

GPT-5.4 looks like a model you want for breadth, not blind faith. For developers, the interesting story is not that it knows a lot, but that calibration still lags behind capability.

• –AA-Omniscience explicitly rewards abstaining when uncertain and punishes confident wrong answers, so this is a benchmark about trustworthiness, not just trivia recall
• –GPT-5.4 (xhigh) posts roughly 50% AA-Omniscience accuracy, which is strong, but its tendency to guess drags down its standing on the benchmark’s reliability framing
• –That makes it a good fit for workflows with retrieval, tool use, or human review, and a riskier fit for unattended knowledge work where a polished wrong answer is costly
• –The benchmark also highlights a broader market split: labs are shipping smarter models faster than they are shipping well-calibrated ones
• –If this pattern holds, the winning product layer will be systems that force verification rather than apps that trust model confidence at face value