OpenAI Introduces GPT 5.2: A Lengthy Context Workhorse For Brokers, Coding And Information Work

OpenAI has simply launched GPT-5.2, its most superior frontier mannequin for skilled work and lengthy working brokers, and is rolling it out throughout ChatGPT and the API.

GPT-5.2 is a household of three variants. In ChatGPT, customers see ChatGPT-5.2 Prompt, Considering and Professional. Within the API, the corresponding fashions are gpt-5.2-chat-latest, gpt-5.2, and gpt-5.2-pro. Prompt targets on a regular basis help and studying, Considering targets complicated multi step work and brokers, and Professional allocates extra compute for onerous technical and analytical duties.

Benchmark profile, from GDPval to SWE Bench

GPT-5.2 Considering is positioned as the principle workhorse for actual world information work. On GDPval, an analysis of effectively specified information duties throughout 44 occupations in 9 giant industries, it beats or ties prime business professionals on 70.9 % of comparisons, whereas producing outputs at greater than 11 occasions the pace and beneath 1 % of the estimated skilled value. For engineering groups this implies the mannequin can reliably generate artifacts corresponding to shows, spreadsheets, schedules, and diagrams given structured directions.

On an inside benchmark of junior funding banking spreadsheet modeling duties, common scores rise from 59.1 % with GPT-5.1 to 68.4 % with GPT-5.2 Considering and 71.7 % with GPT-5.2 Professional. These duties embrace three assertion fashions and leveraged buyout fashions with constraints on formatting and citations, which is consultant of many structured enterprise workflows.

In software program engineering, GPT-5.2 Considering reaches 55.6 % on SWE-Bench Professional and 80.0 % on SWE-bench Verified. SWE-Bench Professional evaluates repository stage patch technology over a number of languages, whereas SWE-bench Verified focuses on Python.

Lengthy context and agentic workflows

Lengthy context is a core design goal. GPT-5.2 Considering units a brand new cutting-edge on OpenAI MRCRv2, a benchmark that inserts a number of similar ‘needle’ queries into lengthy dialogue “haystacks” and measures whether or not the mannequin can reproduce the right reply. It’s the first mannequin reported to achieve close to one hundred pc accuracy on the 4 needle MRCR variant out to 256k tokens.

For workloads that exceed even that context, GPT-5.2 Considering integrates with the Responses /compact endpoint, which performs context compaction to increase the efficient window for instrument heavy, lengthy working jobs. That is related in case you are constructing brokers that iteratively name instruments over many steps and want to keep up state past the uncooked token restrict.

On instrument utilization, GPT-5.2 Considering reaches 98.7 % on Tau2-bench Telecom, a multi flip buyer help benchmark the place the mannequin should orchestrate instrument calls throughout a practical workflow. The official examples from OpenAI launch submit present eventualities like a traveler with a delayed flight, missed connection, misplaced bag and medical seating requirement, the place GPT-5.2 manages rebooking, particular help seating and compensation in a constant sequence whereas GPT-5.1 leaves steps unfinished.

Imaginative and prescient, science and math

Imaginative and prescient high quality additionally strikes up. GPT-5.2 Considering roughly halves error charges on chart reasoning and consumer interface understanding benchmarks like CharXiv Reasoning and ScreenSpot Professional when a Python instrument is enabled. The mannequin reveals improved spatial understanding of photos, for instance when labeling motherboard elements with approximate bounding bins, GPT-5.2 identifies extra areas with tighter placement than GPT-5.1.

For scientific workloads, GPT-5.2 Professional scores 93.2 % and GPT-5.2 Considering 92.4 % on GPQA Diamond, and GPT-5.2 Considering solves 40.3 % of FrontierMath Tier 1 to Tier 3 issues with Python instruments enabled. These benchmarks cowl graduate stage physics, chemistry, biology and skilled arithmetic, and OpenAI highlights early use the place GPT-5.2 Professional contributed to a proof in statistical studying concept beneath human verification.

Comparability Desk

Key Takeaways

1. GPT-5.2 Considering is the brand new default workhorse mannequin: It replaces GPT-5.1 Considering as the principle mannequin for coding, information work and brokers, whereas retaining the identical 400k context and 128k max output, however with clearly greater benchmark efficiency throughout GDPval, SWE-Bench, ARC-AGI and scientific QA.
2. Substantial accuracy leap over GPT-5.1 at comparable scale: On key benchmarks, GPT-5.2 Considering strikes from 50.8 % to 55.6 % on SWE-Bench Professional and from 76.3 % to 80.0 % on SWE-bench Verified, and from 72.8 % to 86.2 % on ARC-AGI-1 and from 17.6 % to 52.9 % on ARC-AGI-2, whereas retaining token limits comparable.
3. GPT-5.2 Professional is focused at excessive finish reasoning and science: GPT-5.2 Professional is a better compute variant that primarily improves onerous reasoning and scientific duties, for instance reaching 93.2 % on GPQA Diamond versus 92.4 % for GPT-5.2 Considering and 88.1 % for GPT-5.1 Considering, and better scores on ARC-AGI tiers.