OpenAI introduces GPT 5.2: a long-term reference workhorse for agents, coding, and knowledge work

OpenAI has just started GPT-5.2Its the most advanced Frontier model for professional work and long-term agents, and it is being introduced in ChatGPTT and API.

GPT-5.2 is a family of three variants. In ChatGPT, users see ChatGPT-5.2 Instant, Thinking, and Pro. In the API, there are related models gpt-5.2-chat-latest, gpt-5.2And gpt-5.2-proTargeting everyday support and accelerated learning, Thinking targets complex multi-step tasks and agents, and Pro allocates more compute for difficult technical and analytical tasks,

Benchmark profiles, from GDPval to SWE bench

GPT-5.2 thinking is deployed as the main workspace for real-world knowledge work. On GDPeval, evaluating well-specified knowledge tasks across 44 businesses across 9 large industries, it outperforms or matches top industry professionals by 70.9 percent, while delivering output at more than 11 times the speed and less than 1 percent of the estimated expert cost. For engineering teams this means that the model can reliably generate artifacts such as presentations, spreadsheets, schedules, and diagrams given structured instructions.

On internal benchmarks of junior investment banking spreadsheet modeling tasks, the average score increased from 59.1 percent with GPT-5.1 to 68.4 percent with GPT-5.2 Thinking and 71.7 percent with GPT-5.2 Pro. These functions include a three statement model and a leveraged buyout model with constraints on formatting and citations, which is representative of many structured enterprise workflows.

In software engineering, GPT-5.2 thinking reaches 55.6 percent on SWE-Bench Pro and 80.0 percent on SWE-Bench Verified. SWE-Bench Pro evaluates repository level patch generation in multiple languages, while SWE-Bench Verified focuses on Python.

Long context and agentic workflow

Long context is a main design goal. GPT-5.2 Thinking establishes the state of the art on OpenAI MRCRv2, a benchmark that inserts several identical ‘needle’ questions into long dialogue “haystacks” and measures whether the model can reproduce the correct answer. This is the first model reported to reach 100 percent accuracy on the 4 needle MRCR version for up to 256k tokens.

For workloads even larger than that, GPT-5.2 integrates with Thinking Reactions /compact Endpoint, which tools perform context compression to expand the effective window for heavy, long-running jobs. This is relevant if you are creating agents that call tools iteratively across multiple steps and need to maintain state beyond the raw token limit.

On tool usage, GPT-5.2 Thinking reaches 98.7 percent on Tau2-Bench Telecom, a multi-turn customer support benchmark where the model must organize tool calls into a realistic workflow. Official examples in the OpenAI release post show scenarios such as delayed flights, missed connections, lost bags and passengers requiring medical seating, where GPT-5.2 handles rebooking, special assistance seating and compensation in a sequential sequence while GPT-5.1 leaves the steps incomplete.

vision, science and mathematics

Vision quality also improves. GPT-5.2 Thinking nearly halves the error rate on user interfaces that understand chart reasoning and benchmarks like CharXiv Reasoning and ScreenSpot Pro when Python tools are enabled. The model shows better spatial understanding of images, for example when motherboard components are labeled with estimated bounding boxes, GPT-5.2 identifies more areas with strict placement than GPT-5.1.

For scientific workloads, GPT-5.2 Pro scores 93.2 percent and GPT-5.2 Thinking scores 92.4 percent on GPQA Diamond, and GPT-5.2 Thinking solves 40.3 percent of FrontierMath Tier 1 to Tier 3 problems when the Python tool is enabled. These benchmarks cover undergraduate level physics, chemistry, biology and specialist mathematics, and highlight an OpenAI initial use where GPT-5.2 Pro contributed to proof in statistical learning theory under human verification.

comparison table

key takeaways

1. GPT-5.2 thinking is the new default workhorse model: It replaces GPT-5.1 Thinking as the main model for coding, knowledge work, and agents, while keeping the same 400k context and 128k maximum output, but with markedly higher benchmark performance in GDPval, SWE-Bench, ARC-AGI, and Scientific QA.
2. Substantial accuracy leap over GPT-5.1 on similar scale: On key benchmarks, GPT-5.2 thinking increases from 50.8 percent to 55.6 percent on SWE-Bench Pro and 76.3 percent to 80.0 percent on SWE-Bench Verified, and from 72.8 percent to 86.2 percent on ARC-AGI-1 and 17.6 percent to 52.9 percent on ARC-AGI-2, while the token limit Is comparable.
3. GPT-5.2 Pro is targeted at high-level logic and science: GPT-5.2 Pro is a higher compute version that primarily improves on difficult reasoning and scientific tasks, for example reaching 93.2 percent on GPQA Diamond, compared to 92.4 percent for GPT-5.2 Thinking and 88.1 percent for GPT-5.1 Thinking, and higher scores on ARC-AGI levels.