AI Explained

GPT 5.2: OpenAI Strikes Back

The video examines OpenAI’s latest release, GPT‑5.2, which OpenAI touts as the first model to reach human‑expert level on the GDPVAL benchmark, beating or tying top professionals on 71% of tasks. The presenter frames the launch as a “luxury Christmas present” for the AI community, while cautioning that many of the headline results may be driven by heavy token‑spending and may be short‑lived. Key insights focus on how benchmark performance is increasingly a function of test‑time compute and token budgets. GPT‑5.2 scores record numbers on GDPVAL, ARK‑AGI 1 (over 90% with extra‑high reasoning effort) and ARK‑AGI 2, yet falls behind Gemini 3 Pro on multimodal segmentation and behind Claude Opus 4.5 on coding and web‑development tasks. The presenter highlights OpenAI’s own admission that more tokens generally yield better scores, and points out the difficulty of fair head‑to‑head comparisons when providers can allocate different compute resources. Notable examples include a football‑season interaction matrix that GPT‑5.2 Pro generated accurately, and a “four‑needle” long‑context test where the model achieved near‑100% recall across 200‑word passages. The video also cites a cheeky comment from former OpenAI staff Logan Kilpatrick that Gemini 3 Pro still leads multimodal understanding, and quotes OpenAI’s Noam Brown on publishing single‑number benchmark results for simplicity despite the need for an x‑axis of token or cost usage. The broader implication is that enterprises must look beyond headline scores and consider token efficiency, pricing, and specific use‑case strengths. GPT‑5.2’s strength lies in long‑context reasoning (up to 400 k tokens) and incremental cost‑effective improvements, but the race for super‑intelligence may continue to be driven by incremental gains rather than a single breakthrough, complicating model selection for businesses.