The wait is over. While industry insiders and previous reports anticipated a launch earlier this week on Tuesday, OpenAI has officially deployed gpt-5.2 today, Thursday, December 11.
Following a rumored “code red” directive from CEO Sam Altman to counter the momentum of Google’s Gemini 3, this release appears to be the structural overhaul reports promised. Moving beyond simple chatter, OpenAI positions GPT-5.2 not just as a chatbot update, but as a “frontier model” specifically engineered for agentic workflows and complex professional tasks.
The release introduces a tiered model family—Instant, Thinking, and Pro—available immediately in ChatGPT and via API.
GPT-5.2 all new features and benchmarks
The headline feature of this release is OpenAI’s performance on GDPval, a new benchmark evaluating 44 real-world professional tasks ranging from spreadsheet creation to legal drafting.
According to OpenAI’s internal data, the gpt-5.2 “Thinking” model is their first to achieve expert-level scores, reportedly beating or tying human industry professionals in 70.9% of comparisons. This is a massive leap from GPT-5.1 Thinking, which only held a 38.8% win/tie rate in the same tests.
Three tiers for different needs
OpenAI has split the GPT-5.2 release into three distinct classes to balance speed, reasoning depth, and cost:
-
GPT-5.2 Instant: Designed for speed. It handles info-seeking, how-tos, and technical writing with low latency.
-
GPT-5.2 Thinking: The “workhorse” for professionals. It features advanced tool-calling and deep reasoning, excelling at spreadsheet formatting, slideshow creation, and long-context analysis.
-
GPT-5.2 Pro: The highest-end model. It is slower and more expensive but designed for critical tasks where accuracy is paramount. It is currently the only model to cross the 90% threshold on the ARC-AGI-1 reasoning benchmark.
Coding and agents
For developers, the gpt-5.2 API release targets the growing market of AI agents. OpenAI claims the model is significantly better at “long-horizon” tasks—complex workflows that require maintaining context over many steps.
-
Coding: On SWE-Bench Pro, a rigorous software engineering benchmark, GPT-5.2 Thinking hit a score of 55.6%, setting a new state-of-the-art (SOTA).
-
Math: On AIME 2025 (competition math), the model achieved a 100% score without tools.
-
Reliability: The company reports a 30% reduction in hallucinations compared to GPT-5.1, a critical improvement for enterprise adoption.
Availability and pricing
The models are rolling out to ChatGPT Plus, Team, and Enterprise users starting today. For API developers, the pricing reflects the model’s increased capability, particularly for the Pro tier.
| Model | Input Cost (per 1M tokens) | Output Cost (per 1M tokens) |
| gpt-5.2 (Thinking/Instant) | $1.75 | $14.00 |
| gpt-5.2-pro | $21.00 | $168.00 |
| gpt-5.1 (Legacy) | $1.25 | $10.00 |
While gpt-5.2 is priced higher than its predecessor, OpenAI argues that its efficiency in “one-shotting” complex tasks makes it cheaper in practice than chaining multiple prompts with weaker models.
