GPT-5.2

Overview

GPT-5.2 is OpenAI's current flagship model, released December 10, 2025. It represents a major step forward for professional knowledge work — benchmarks show it matching or outperforming domain experts across 44 occupations on OpenAI's GDPval evaluation, with judges rating it better than top human professionals on 70.9% of comparisons in the Thinking variant. It also crossed the 90% threshold on ARC-AGI-1, a benchmark specifically designed to resist AI systems that rely purely on pattern matching over genuine reasoning.

GPT-5.2 is available in three modes — Instant, Thinking, and Pro — that let developers trade off between speed, depth of reasoning, and cost depending on the task. This tiered structure makes it practical across a wide range of use cases, from real-time applications needing fast responses to deep research tasks that benefit from extended reasoning chains.

Specifications

• Developer: OpenAI
• Model String: gpt-5.2-2025-12-10
• Release Date: December 10, 2025
• Type: Large Language Model (LLM), Multimodal (text + vision)
• Context Window: 400,000 tokens
• Max Output: 128,000 tokens
• Access: OpenAI API, ChatGPT (Plus/Team/Enterprise/Pro), Azure OpenAI Service
• Pricing:
  • Base: $1.75 per million input tokens / $14.00 per million output tokens
  • Pro: $21.00 per million input tokens / $168.00 per million output tokens

Capabilities

Professional Knowledge Work: GPT-5.2 Thinking beats or ties top industry professionals on 70.9% of comparisons on GDPval tasks spanning 44 occupations — the strongest performance on human professional benchmarks of any model to date.

Advanced Reasoning: GPT-5.2 Thinking sets a new state of the art in long-context reasoning, achieving near-100% accuracy on the 4-needle MRCR variant out to 256K tokens — a significant long-context comprehension milestone.

ARC-AGI Performance: First OpenAI model to cross 90% on ARC-AGI-1, a benchmark designed to test genuine reasoning and adaptability rather than memorized patterns.

Coding: Strong performance on real-world software engineering tasks. Accompanied by GPT-5.2 Codex (released January 2026), a specialized coding variant.

Multimodal: Handles text, images, and documents. Supports vision input for document analysis, chart reading, and image-based tasks.

Three Modes:

• Instant — Fastest, for latency-sensitive applications
• Thinking — Extended reasoning chains for complex analysis; best professional-grade output
• Pro — Maximum capability and depth; highest cost

Limitations

At $1.75–$21.00 per million input tokens (depending on variant), GPT-5.2 is more expensive than mid-tier alternatives from Anthropic and Google for equivalent tasks. The Pro variant in particular ($21/$168) is aimed at scenarios where cost is secondary to maximum capability. For high-volume production workloads, GPT-4.1 or o4-mini offer better economics.

Recent Developments

• December 10, 2025 Launch: Released as OpenAI's most capable model, with a focus on professional knowledge work and agentic tasks.
• GPT-5.2 Codex (January 14, 2026): A coding-specialized variant followed within five weeks, extending GPT-5.2's strengths to software engineering workflows.
• ARC-AGI-1 Milestone: Crossing 90% on ARC-AGI-1 reignited debate about how to measure genuine AI reasoning, with ARC-AGI-2 already being positioned as the next meaningful benchmark.
• February 2026 Context: Released alongside intense competition from Anthropic's Opus 4.6 and Google DeepMind's Gemini 2.5, with OpenAI positioning GPT-5.2 as the leader for professional knowledge work specifically.

Last Updated

February 26, 2026