## Friday, December 12, 2025 # GPT-5.2 RAG Performance: We Tested It Umida Muratbekova We plugged [GPT-5.2](/content/llms/gpt-52/index.html) into the same RAG benchmark we use for our LLM leaderboard and compared it against **9 other frontier models**, including [GPT-5.1](/content/llms/gpt-51/index.html), [Claude Opus 4.5](/content/llms/claude-opus-45/index.html), [Grok-4-Fast](/content/llms/grok-4-fast/index.html), [Gemini 3 Pro](/content/llms/gemini-3-pro-preview/index.html), and the strongest open-source options. All models went through the exact same RAG stack across three workloads: - short factual Q&A - long-form reasoning and synthesis - scientific claim verification In this setup, [GPT-5.2](/content/llms/gpt-52/index.html) lands in the **top cluster**, but behaves differently from [GPT-5.1](/content/llms/gpt-51/index.html) and the other leaders. It writes much shorter answers, uses far fewer tokens, and is especially strong when it has to read evidence and decide whether a claim holds. ## How We Tested Same setup as in our main RAG benchmark: - Retrieval: [text-embedding-3-large](/content/embeddings/openai-text-embedding-3-large/index.html) + [zerank-1](/content/rerankers/zerank-1/index.html), top-15 documents. - Generation: 10 frontier LLMs answering the _same_ questions with the _same_ retrieved context. - Judging: GPT-5, comparing answers in pairs. - Scoring: pairwise wins → ELO; plus per-answer scores on correctness, faithfulness, grounding, relevance, and completeness. ## What We Found About GPT-5.2 A few patterns stand out when you look at [GPT-5.2](/content/llms/gpt-52/index.html) on the leaderboard: - **Overall position:** [GPT-5.2](/content/llms/gpt-52/index.html) sits in the top group, ranked #4 with 1588 ELO. [GPT-5.1](/content/llms/gpt-51/index.html) is #1 at 1711 ELO. - **Token usage:** [GPT-5.2](/content/llms/gpt-52/index.html) uses roughly 3× fewer tokens per answer on average (247 vs 796 for [GPT-5.1](/content/llms/gpt-51/index.html)). - **Scientific reasoning:** On our scientific claim verification tasks, [GPT-5.2](/content/llms/gpt-52/index.html) is the top model, with 1605 ELO, about +17 ELO ahead of [GPT-5.1](/content/llms/gpt-51/index.html) and ahead of every other model we tested. - **Stability across domains:** [GPT-5.2](/content/llms/gpt-52/index.html) is much more stable across the three RAG workloads, with a coefficient of variation of around 0.8% (compared to 7.3% for [GPT-5.1](/content/llms/gpt-51/index.html)). ## 1. Shorter Answers, Similar Quality Across all three workloads, [GPT-5.2](/content/llms/gpt-52/index.html)’s answers are much shorter than [GPT-5.1](/content/llms/gpt-51/index.html)’s. - [GPT-5.1](/content/llms/gpt-51/index.html): ~ **800 tokens** per answer - [GPT-5.2](/content/llms/gpt-52/index.html): ~ **250 tokens** per answer So roughly a **70% drop in token usage**. Even with that gap, the judge scores on correctness, faithfulness, and grounding usually stay close between the two models. On many questions, both are rated highly; [GPT-5.2](/content/llms/gpt-52/index.html) just spends fewer tokens getting there. It removes restatements and side explanations and goes straight to: answer + key evidence. For chat-style products, this difference is mostly about tone. For high-throughput or cost-sensitive RAG systems, cutting answer length by this much has a very direct impact on latency and API cost. ## 2. Where GPT-5.2 Leads: Scientific Claim Verification On scientific claim verification, [GPT-5.2](/content/llms/gpt-52/index.html) is the best model we tested. These queries look like: > Here is a claim and one or two evidence passages. > > Is the claim supported, contradicted, or unclear? Good answers are short and anchored to the text: a clear stance and a pointer to the exact sentence or result. In this setting, [GPT-5.2](/content/llms/gpt-52/index.html) reaches **1605 ELO**, slightly ahead of [GPT-5.1](/content/llms/gpt-51/index.html) and above every other model. In concrete examples, both often choose the same label and cite the same experiment. The difference is how they say it: [GPT-5.1](/content/llms/gpt-51/index.html) adds more narrative and framing; [GPT-5.2](/content/llms/gpt-52/index.html) keeps only what is needed to justify the decision. For scientific and technical RAG—where users move through many such claims—this compact, evidence-first style is closer to what the task actually demands. ## 3. Consistency Across MSMARCO, PG Essays, and SciFact [GPT-5.2](/content/llms/gpt-52/index.html) is also more **stable** across the different workloads. If we look at its ELO scores per workload and compute the coefficient of variation: - [GPT-5.2](/content/llms/gpt-52/index.html): ~ **0.8%** - [GPT-5.1](/content/llms/gpt-51/index.html): ~ **7.3%** [GPT-5.2](/content/llms/gpt-52/index.html) stays close to its baseline level whether the query is a short factual question, a long reasoning prompt, or a scientific claim. [GPT-5.1](/content/llms/gpt-51/index.html) (and some of the other leaders) hit slightly higher peaks in some areas, but move more between them. Put together with the token findings, [GPT-5.2](/content/llms/gpt-52/index.html) looks less like a “general assistant” and more like a **compact specialist**: high RAG quality, relatively uniform across workloads, and a much smaller token footprint per answer. ## When is GPT-5.2 a Good Fit? This benchmark suggests [GPT-5.2](/content/llms/gpt-52/index.html) is a good match when: - you are serving scientific or technical users who mainly care about the conclusion and the supporting evidence - cost and latency are real constraints, so large cuts in token usage matter - your RAG traffic mixes factual, reasoning, and scientific queries, and you want steady behaviour across them - your product benefits more from concise, information-dense outputs than from long, conversational explanations In those cases, [GPT-5.2](/content/llms/gpt-52/index.html) behaves in line with the goal: it reads the retrieved documents, takes a clear stance, cites the key lines, and stops. If you want to see how [GPT-5.2](/content/llms/gpt-52/index.html) compares to [GPT-5.1](/content/llms/gpt-51/index.html), [Grok-4-Fast](/content/llms/grok-4-fast/index.html), [Claude Opus 4.5](/content/llms/claude-opus-45/index.html), [Gemini 3 Pro](/content/llms/gemini-3-pro-preview/index.html), [GLM](/content/llms/glm-46/index.html), and the strongest open-source models, the full breakdown is available on our [LLM leaderboard](/content/llms/index.html).