OpenAI has simply launched GPT-5.2, its most superior frontier mannequin for skilled work and lengthy operating 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 Immediate, Pondering and Professional. Within the API, the corresponding fashions are gpt-5.2-chat-latest, gpt-5.2, and gpt-5.2-pro. Immediate targets on a regular basis help and studying, Pondering targets advanced multi step work and brokers, and Professional allocates extra compute for laborious technical and analytical duties.
Benchmark profile, from GDPval to SWE Bench
GPT-5.2 Pondering is positioned as the principle workhorse for actual world data work. On GDPval, an analysis of nicely specified data duties throughout 44 occupations in 9 giant industries, it beats or ties prime trade professionals on 70.9 p.c of comparisons, whereas producing outputs at greater than 11 instances the velocity and beneath 1 p.c of the estimated knowledgeable value. For engineering groups this implies the mannequin can reliably generate artifacts similar to displays, spreadsheets, schedules, and diagrams given structured directions.
On an inner benchmark of junior funding banking spreadsheet modeling duties, common scores rise from 59.1 p.c with GPT-5.1 to 68.4 p.c with GPT-5.2 Pondering and 71.7 p.c 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 Pondering reaches 55.6 p.c on SWE-Bench Professional and 80.0 p.c on SWE-bench Verified. SWE-Bench Professional evaluates repository degree 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 Pondering units a brand new cutting-edge on OpenAI MRCRv2, a benchmark that inserts a number of equivalent ‘needle’ queries into lengthy dialogue “haystacks” and measures whether or not the mannequin can reproduce the proper 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 Pondering integrates with the Responses /compact endpoint, which performs context compaction to increase the efficient window for instrument heavy, lengthy operating jobs. That is related in case you are constructing brokers that iteratively name instruments over many steps and want to take care of state past the uncooked token restrict.
On instrument utilization, GPT-5.2 Pondering reaches 98.7 p.c on Tau2-bench Telecom, a multi flip buyer help benchmark the place the mannequin should orchestrate instrument calls throughout a sensible workflow. The official examples from OpenAI launch publish 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 Pondering roughly halves error charges on chart reasoning and person interface understanding benchmarks like CharXiv Reasoning and ScreenSpot Professional when a Python instrument is enabled. The mannequin reveals improved spatial understanding of photographs, for instance when labeling motherboard parts 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 p.c and GPT-5.2 Pondering 92.4 p.c on GPQA Diamond, and GPT-5.2 Pondering solves 40.3 p.c of FrontierMath Tier 1 to Tier 3 issues with Python instruments enabled. These benchmarks cowl graduate degree physics, chemistry, biology and knowledgeable 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
| Mannequin | Major positioning | Context window / max output | Information cutoff | Notable benchmarks (Pondering / Professional vs GPT-5.1 Pondering) |
|---|---|---|---|---|
| GPT-5.1 | Flagship mannequin for coding and agentic duties with configurable reasoning effort | 400,000 tokens context, 128,000 max output | 2024-09-30 | SWE-Bench Professional 50.8 p.c, SWE-bench Verified 76.3 p.c, ARC-AGI-1 72.8 p.c, ARC-AGI-2 17.6 p.c |
| GPT-5.2 (Pondering) | New flagship mannequin for coding and agentic duties throughout industries and for lengthy operating brokers | 400,000 tokens context, 128,000 max output | 2025-08-31 | GDPval wins or ties 70.9 p.c vs trade professionals, SWE-Bench Professional 55.6 p.c, SWE-bench Verified 80.0 p.c, ARC-AGI-1 86.2 p.c, ARC-AGI-2 52.9 p.c |
| GPT-5.2 Professional | Larger compute model of GPT-5.2 for the toughest reasoning and scientific workloads, produces smarter and extra exact responses | 400,000 tokens context, 128,000 max output | 2025-08-31 | GPQA Diamond 93.2 p.c vs 92.4 p.c for GPT-5.2 Pondering and 88.1 p.c for GPT-5.1 Pondering, ARC-AGI-1 90.5 p.c and ARC-AGI-2 54.2 p.c |
Key Takeaways
- GPT-5.2 Pondering is the brand new default workhorse mannequin: It replaces GPT-5.1 Pondering as the principle mannequin for coding, data work and brokers, whereas conserving the identical 400k context and 128k max output, however with clearly increased benchmark efficiency throughout GDPval, SWE-Bench, ARC-AGI and scientific QA.
- Substantial accuracy leap over GPT-5.1 at comparable scale: On key benchmarks, GPT-5.2 Pondering strikes from 50.8 p.c to 55.6 p.c on SWE-Bench Professional and from 76.3 p.c to 80.0 p.c on SWE-bench Verified, and from 72.8 p.c to 86.2 p.c on ARC-AGI-1 and from 17.6 p.c to 52.9 p.c on ARC-AGI-2, whereas conserving token limits comparable.
- GPT-5.2 Professional is focused at excessive finish reasoning and science: GPT-5.2 Professional is a better compute variant that primarily improves laborious reasoning and scientific duties, for instance reaching 93.2 p.c on GPQA Diamond versus 92.4 p.c for GPT-5.2 Pondering and 88.1 p.c for GPT-5.1 Pondering, and better scores on ARC-AGI tiers.
Asif Razzaq is the CEO of Marktechpost Media Inc.. As a visionary entrepreneur and engineer, Asif is dedicated to harnessing the potential of Synthetic Intelligence for social good. His most up-to-date endeavor is the launch of an Synthetic Intelligence Media Platform, Marktechpost, which stands out for its in-depth protection of machine studying and deep studying information that’s each technically sound and simply comprehensible by a large viewers. The platform boasts of over 2 million month-to-month views, illustrating its recognition amongst audiences.
