XNews

OpenAI launches GPT-5.2 (benchmarks & capabilities)

OpenAI launches GPT-5.2 (benchmarks & capabilities)

OpenAI released GPT-5.2, a multi-edition model suite aimed at professional knowledge work. - Release: GPT-5.2 announced with three editions: Instant, Thinking, Pro. - Benchmarks: SWE-Bench 55.6% (from 50.8%). - Benchmarks: GPQA Diamond 92.4% (from 88.1%). - Benchmarks: CharXiv scientific reasoning 82.1% (from 67.0%). - Benchmarks: FrontierMath overall 40.3% (from 31.0%); Tier 4 14.6%. - Benchmarks: AIME 2025 perfect 100% (from 94.0%). - Benchmarks: ARC-AGI1 86.2% (from 72.8%). - Benchmarks: ARC-AGI2 52.9% (from 17.6%). - GDPval: knowledge-work tasks 70.9% (from 38.8%), covering 44 professions and 1,320 tasks. - Long context: near-100% accuracy on 256k-token tests. - Positioning: focused on “professional knowledge work” and long-range agents. - Versions purpose: Instant for fast tasks, Thinking for deep multi-step reasoning, Pro for highest compute use cases. - Product strategy: priced ~40% higher; rollout to paid users starting this week. - Competitive context: framed as response to Gemini 3 Pro and Claude Opus 4.5; cited internal “Code Red” effort. - Tooling: improved tool-calling and multi-step execution; claimed 30% lower error on complex projects vs prior. - Features: enterprise push, infrastructure spending commitment, and planned “adult mode” with age-prediction rollout and adult-mode in Q1 next year. - Language quality: demonstrated high-quality literary/style imitation in sample outputs.

Key insights

GPT-5.2 introduced as a multi-edition suite targeting knowledge work

OpenAI announced GPT-5.2 delivered as three distinct editions—Instant for fast responses, Thinking for deep multi-step reasoning, and Pro for highest-compute research use. The release positions GPT-5.2 as a specialist ‘expert colleague’ aimed at professional knowledge work, shifting from general chat focus to reliability on long multi-step tasks and complex domain outputs such as spreadsheets, financial models, and research synthesis. The company has prioritized enterprise readiness and monetization, rolling the model to paid customers this week and raising prices by ~40%.

Benchmark improvements show large gains across scientific and reasoning tests

Multiple benchmark results were reported for GPT-5.2 Thinking: SWE-Bench software engineering 55.6% (from 50.8%), GPQA Diamond 92.4% (from 88.1%), CharXiv scientific reasoning 82.1% (from 67.0%), FrontierMath overall 40.3% (from 31.0%) with Tier 4 at 14.6%, AIME 2025 perfect 100% (from 94.0%), ARC-AGI1 86.2% (from 72.8%) and ARC-AGI2 52.9% (from 17.6%). These figures were presented as evidence of major capability jumps over GPT-5.1 and as competitive gains versus Claude Opus 4.5 and Gemini 3 Pro.

GDPval metric shows higher knowledge-work parity with humans

OpenAI cited a new GDPval benchmark covering 1,320 tasks across 44 professions; GPT-5.2 Thinking scored 70.9% versus prior ~38.8%. OpenAI presented GDPval as a measure of real task output (deliverables like spreadsheets, slide decks, schedules) and used it to argue GPT-5.2 can match or exceed human experts on a large share of knowledge-work tasks.

Long-context and mathematical performance emphasized

OpenAI reported near-100% precision on a 256k-token long-context ‘needle-in-haystack’ test and highlighted perfect AIME 2025 results without external tools. The company emphasized improved long-range coherence and advanced math/reasoning performance as core upgrades in the Thinking edition.

Strategic response to Google and Anthropic; internal Code Red mobilization

OpenAI framed GPT-5.2 as part of a recovery effort following competitive pressure from Google’s Gemini 3 and Anthropic’s releases. Sam Altman declared a ‘Code Red’ mobilization earlier, directing resources to accelerate ChatGPT improvements; OpenAI expects to exit Code Red by January and aims to regain market momentum.

Enterprise features, pricing, and safety controls included

OpenAI emphasized enterprise adoption: higher pricing (~40% increase), plans to offer an age-prediction system to gate content, a planned ‘adult mode’ in Q1 with age verification, and commitments to safety and IP protections. The release includes product integration focus—improved tool calls, multi-step execution, and enterprise efficiency claims (time and cost savings).

Source posts