GPT-5.4 has strong public evidence for coding and agent use, though it is not a clear category leader. Terminal-Bench reports 77.3% with Codex CLI on Terminal-Bench 2.1, and a third-party LiveCodeBench reporting page lists Codex GPT-5.4 at 91.8% pass@1. Its 1.1M-token context and Artificial Analysis Intelligence Index of 51 support a high technical score. It remains below the published GPT-5.5 and Claude Opus 4.8 anchors in technical and agentic work, whose evaluations carry stronger top-end evidence. The coding score is moderated by DeepSWE: gpt-5.4 xhigh scored 52% ±2% and ranked 12th of 18 displayed configurations. Current Text Arena placement (#12 overall, #8 Expert, #12 Hard Prompts) supports solid but not leading reasoning/public-preference performance; this is stronger than lower-reasoning anchors such as GPT-5.6 Terra, but below GPT-5.5, GPT-5.6 Sol, and Opus 4.8. Official documentation supports API, tool-use, and safety-material availability, while the supplied evidence does not substantiate detailed modality behavior beyond official capability claims. At $2.50/M input and $15/M output, GPT-5.4 is materially less cost-effective than DeepSeek-V3.2 or Gemma 3, despite the large context window. Reasoning-mode latency is a constraint: Artificial Analysis reports 121.68s TTFT, versus 0.80s for non-reasoning; output throughput is high after generation begins. Pricing documentation, model docs, system-card material, and Codex integration make the developer experience strong. Evidence quality is good but limited by missing directly evidenced SWE-bench results and the discrepancy between official LiveCodeBench absence and a third-party Codex listing.
