Simon Willison's Weblog
•December 11, 2025GPT-5.2
Quick Overview
OpenAI has released GPT-5.2 and GPT-5.2 Pro, significantly enhancing their model series for professional knowledge work and vision tasks.
- Improved Performance: Achieves 70.9% on GDPval "Knowledge work tasks" and 52.9% on ARC-AGI-2 benchmarks.
- Extended Knowledge Cut-off: Now August 31st, 2025, a significant update from previous versions.
- Enhanced Vision Capabilities: Cuts error rates in half for chart reasoning and software interface understanding.
- Cost Increase: GPT-5.2 is 1.4x the cost of GPT-5.1, and GPT-5.2 Pro is considerably more expensive.
- Efficiency Gains: GPT-5.2 Pro demonstrates a ~390x efficiency improvement in one year for specific tasks.
Key Points
Launch Context and Purpose
- OpenAI reportedly declared a 'code red' on December 1st due to increased competition from models like Google's Gemini 3.
- Less than two weeks later, OpenAI announced GPT-5.2, marketing it as 'the most capable model series yet for professional knowledge work'.
- This rapid release suggests a strategic response to competitive pressures in the AI landscape.
Key Features and Variants
- GPT-5.2 is available in two variants: GPT-5.2 and GPT-5.2 Pro, with no 'Mini' version yet.
- The models are accessible via OpenAI's UI with 'instant' and 'thinking' modes, and through new API models: 'gpt-5.2', 'gpt-5.2-chat-latest', and 'gpt-5.2-pro'.
- Both 5.2 models boast a 400,000 token context window and 128,000 max output tokens, consistent with previous 5.x versions.
- A significant update is the knowledge cut-off date, now extended to August 31st, 2025.
Performance and Cost Metrics
- GPT-5.2 achieved a 70.9% score on OpenAI's GDPval 'Knowledge work tasks' benchmark, significantly up from GPT-5's 38.8%.
- It scored 52.9% on ARC-AGI-2, an increase from GPT-5.1 Thinking's 17.6%.
- Pricing for GPT-5.2 is 1.4x the cost of GPT 5.1, at $1.75/million input and $14/million output.
- GPT-5.2 Pro is considerably more expensive at $21.00/million input and $168.00/million output.
- An 'ARC Prize' verification noted a ~390X efficiency improvement for GPT-5.2 Pro over a year, achieving 90.5% on ARC-AGI-1 at $11.64/task compared to an earlier model's 88% at $4.5k/task.
Enhanced Capabilities
- Vision capabilities: GPT-5.2 Thinking is described as OpenAI's 'strongest vision model yet', roughly halving error rates in chart reasoning and software interface understanding.
- Prompting Guide: OpenAI has published a new 'GPT-5.2 Prompting Guide' to assist users.
- Response Compaction: For long-running, tool-heavy workflows, GPT-5.2 with Reasoning supports a new server-side API, '/responses/compact', which performs 'loss-aware compression' of prior conversation state to extend context without hitting limits.
Outline
No outline available
AI saves you up to 6 minutes
