GPT-5.2 at Launch: Benchmark Results, Demos and API Pricing

要約

30 秒でわかる

In this short December 2025 video, Matthew Berman presents OpenAI’s launch figures for GPT-5.2 and examples of spreadsheets, software and image understanding. The comparisons are from OpenAI’s selected tests and demonstrations rather than an independent assessment. He also reviews launch-era API prices and ChatGPT rollout plans. The GDPval comparison shown is with GPT-5, despite his description as a comparison with GPT-5.1.

1Knowledge-work comparisonOpenAI’s GDPval table reports 70.9% wins or ties for GPT-5.2 Thinking versus 38.8% for GPT-5.
2Coding and mathThe launch chart lists 55.6% on SWE-bench Pro and 100% on AIME 2025 for GPT-5.2 Thinking.
3Abstract reasoningOpenAI reports 52.9% on ARC-AGI-2 Verified for GPT-5.2 Thinking, compared with 17.6% for GPT-5.1 Thinking.
4Historical API pricingThe launch chart lists $1.75 per million input tokens and $14 per million output tokens for standard GPT-5.2.

ポイント

OpenAI reported GPT-5.2 Thinking at 70.9% wins or ties on GDPval versus 38.8% for GPT-5, not GPT-5.1.
The launch table reported 55.6% on SWE-bench Pro and 100% on AIME 2025 for GPT-5.2 Thinking under its stated conditions.
OpenAI reported 52.9% on ARC-AGI-2 Verified for GPT-5.2 Thinking versus 17.6% for GPT-5.1 Thinking.
The spreadsheet, software and motherboard examples are selected demonstrations and do not establish universal performance.
At launch, standard GPT-5.2 API input and output prices shown were higher than those for GPT-5.1.

トピックOpenAI Benchmarks AI cost & performance コーディングエージェント画像と動画

全文

チャプターごとに整えた全文を読む

元の文字起こしと対照

文字起こしのみ

OpenAI’s benchmark claims

Berman highlights launch results for knowledge work, coding, mathematics and abstract reasoning. The displayed GDPval baseline is GPT-5, despite his reference to GPT-5.1.

スライド 1 / 4Launch benchmark tableOpenAI’s table labels the GDPval baseline as GPT-5 and lists selected coding and mathematics results.

In December 2025, Matthew Berman reviews OpenAI’s introduction of GPT-5.2, highlighting its reported benchmark improvements. The launch chart shows GPT-5.2 Thinking winning or tying human professionals on 70.9% of GDPval tasks versus 38.8% for GPT-5; Berman describes the baseline as GPT-5.1, but the displayed table identifies GPT-5.

スライド 2 / 4ARC-AGI-2 comparisonThe launch table reports 52.9% for GPT-5.2 Thinking and 17.6% for GPT-5.1 Thinking.

OpenAI’s table reports GPT-5.2 Thinking at 55.6% on SWE-bench Pro versus 50.8% for GPT-5.1 Thinking, and 100% versus 94% on AIME 2025 without tools. On ARC-AGI-2 Verified, the reported scores are 52.9% and 17.6% respectively. These are provider-reported research evaluations under specified conditions, not proof of general intelligence or a universal ranking.

Selected demos and pricing

The video shows spreadsheets, a project-management example, an ocean-wave simulation and motherboard labels. Berman closes with the initial rollout and API prices.

Berman compares selected examples of a spreadsheet and a project-management output, saying the newer model produces a more polished result. He then shows an ocean-wave simulation with controls for wind speed, wave height and lighting. These examples illustrate outputs shown at launch rather than a systematic comparison of every task.

For image understanding, he shows a motherboard-labeling comparison and says GPT-5.2 identifies more components with more precise markings. The video presents this as one visual example, not an independently verified error rate across images.

スライド 4 / 4Launch-era API pricesThe table compares input, cached-input and output pricing for the model versions shown.

Berman says GPT-5.2 Instant, Thinking and Pro began rolling out to paid ChatGPT plans at launch. The displayed API table lists standard GPT-5.2 at $1.75 per million input tokens and $14 per million output tokens, compared with $1.25 and $10 for GPT-5.1. These are launch-era prices and rollout details.