Official data positions GPT-5.2 Thinking/Pro as models for complex reasoning, coding, long documents, and tool tasks, but these benchmarks cannot be directly treated as Chat experience scores for gpt-5.2-chat-latest.
Release versions: GPT-5.2 Instant, GPT-5.2 Thinking, and GPT-5.2 Pro; API-related names include gpt-5.2, gpt-5.2-chat-latest, and gpt-5.2-pro.
Inputs/modalities: The official release covers text, images, long context, and tool/Agent scenarios; the specific harness for each benchmark is described separately in the release notes.
Configuration: Thinking/Pro use the official evaluation settings; the report does not disclose the complete original inputs, random seeds, or all tool traces for each table.
The official release examples include a workforce planning task and a single-HTML waveform simulator prompt; the release notes also position Instant for everyday work, Thinking for coding, long documents, files, and logic, and Pro for the highest quality requirements.
Representative GPT-5.2 Thinking scores reported officially include: GDPval 70.9%, SWE-Pro 55.6%, SWE Verified 80.0%, GPQA Diamond 92.4%, CharXiv Python 88.7%, HMMT 99.4%, FrontierMath T1–3 40.3%, ARC-AGI-1 86.2%, ARC-AGI-2 52.9%, Tau2 Telecom 98.7%, BrowseComp 65.8%, MCP Atlas 60.6%, and Toolathlon 46.3%. On the long-context MRCR v2 benchmark, the score was 98.2% for 4k–8k 8-needle and 77.0% for 128k–256k.
The official report also says that, compared with GPT-5.1, the frequency of responses containing errors fell by 38% on de-identified ChatGPT queries for GPT-5.2 Thinking; the error detection was performed by a model, and the official report cautions that the detector may be wrong.
For scenarios involving “complex problems + sufficient reasoning/tool budget,” the official data supports using GPT-5.2 Thinking/Pro as strong baselines. If the goal is low-latency conversation aligned with ChatGPT, gpt-5.2-chat-latest should be evaluated separately; Thinking scores cannot simply be transferred to it.
This is a vendor-reported benchmark, and it does not disclose all original samples, prompts, failure cases, random seeds, or complete cost traces.
Different variants, reasoning effort, tools, and server-side paths are mixed together in the release notes; they must be separated by task rather than reduced to a single overall ranking.
The Chat version's model page was marked deprecated on 2026-08-18, and its context/output limits on Chat Completions differ from those of gpt-5.2.
The 38% is a relative change in the frequency of detected errors, not an absolute accuracy improvement across all user tasks.
Specify whether to evaluate gpt-5.2-chat-latest or gpt-5.2/Thinking, and fix the snapshot, API, reasoning, and verbosity.
Run Chat and Thinking separately on the same task set; record inputs, outputs, tool traces, tokens, latency, and error types.
For tasks requiring fresh facts, provide the same search tool consistently; for long-context tests, record the 4k–8k and 128k–256k ranges separately at minimum.
Use the official tables only as a comparison, and do not present scores from an undisclosed harness as an independent reproduction.
GPT-5.2 Chat
