OpenAI Fires Back With GPT-5.2, Tripling Its ARC-AGI-2 Score in Four Weeks
The December release targets professional knowledge work, posts 70.9% on GDPval and costs more than GPT-5.1.
Key takeaways
- GPT-5.2 Thinking scored 52.9% on ARC-AGI-2, up from 17.6% for GPT-5.1 Thinking and above Gemini 3 Pro's 31.1%.
- OpenAI reports 70.9% on GDPval, and a drop in responses containing errors from 8.8% to 6.2%.
- API pricing rose to $1.75 input and $14 output per million tokens from $1.25 and $10.
OpenAI released GPT-5.2 on December 11, 2025, four weeks after GPT-5.1 and three weeks after Gemini 3 Pro took the leaderboards. The model arrives in Instant, Thinking and Pro variants and leans hard on one claim: it is the best model yet for professional knowledge work.
The scorecard
The biggest jump is on ARC-AGI-2, a test of novel abstract reasoning that resists memorization. GPT-5.2 Thinking scored 52.9%, versus 17.6% for GPT-5.1 Thinking and 31.1% for Gemini 3 Pro. Another source lists Claude Opus 4.5 at 37.6%. On GDPval, which measures work product across dozens of occupations, OpenAI reports 70.9% against 38.8% for GPT-5 Thinking.
Other numbers moved less dramatically. SWE-bench Verified rose to 80.0% from 76.3%, nearly level with the best published results from Anthropic. GPQA Diamond reached 92.4%, AIME 2025 hit 100% without tools, and ScreenSpot-Pro, which tests reading user interfaces, climbed from 64.2% to 86.3%. Gemini 3 Pro still led on Video-MMMU (87.6% to 85.9%) and multilingual MMMLU (91.8% to 89.6%), so the picture is a lead on most, not all, of the tests The Decoder compared.
Three weeks, three flagships
The timeline shows how compressed the frontier has become. OpenAI shipped GPT-5.1 on November 12, Google followed with Gemini 3 Pro on November 18, Anthropic with Opus 4.5 on November 24, and OpenAI answered on December 11. No lab held the top benchmark position for longer than a few weeks.
Price goes up, not down
Where Anthropic cut prices, OpenAI raised them. GPT-5.2 costs $1.75 per million input tokens and $14 per million output, up from $1.25 and $10 for GPT-5.1. GPT-5.2 Pro is far dearer at $21 and $168. That still undercuts Opus 4.5's $5 and $25 and sits close to Gemini 3 Pro's $2 and $12. OpenAI is betting that a clearer capability lead justifies a higher base rate, and it also says responses contain fewer errors: 6.2% of anonymized ChatGPT answers had at least one mistake, down from 8.8% on GPT-5.1 Thinking.
Why ARC-AGI-2 and GDPval are the headline tests
ARC-AGI-2 is made of small visual puzzles in which a model must infer a rule from a handful of examples and apply it to a new grid. The tasks are easy for most people and hard for models that rely on memorised patterns, which is why the benchmark became a favourite measure of fluid reasoning. Moving from 17.6% to 52.9% between two models released four weeks apart is a large leap, and it likely reflects heavier test-time reasoning rather than a new base model. The ARC Prize organisers also track cost per task, and a score bought with a very large compute budget tells a different story from one achieved cheaply.
GDPval takes the opposite approach. Instead of abstract puzzles, it asks models to produce deliverables such as spreadsheets, slide decks and briefs across a set of occupations, then has professionals compare the results with human work. OpenAI reports 70.9% for GPT-5.2 Thinking versus 38.8% for GPT-5 Thinking, and says investment-banking tasks improved from an average of 59.1% to 68.4%. Because OpenAI designed and graded the benchmark, it is the clearest example of a vendor-owned test, and readers should wait for outside replication.
The error-rate figure is arguably the most practical. OpenAI said that on a sample of anonymised ChatGPT requests, 6.2% of GPT-5.2 Thinking responses had at least one error, against 8.8% for GPT-5.1 Thinking. A drop of 2.6 points means roughly three fewer flawed answers in every hundred, which matters more for office workflows than a leaderboard position. It also shows how far models still are from reliable: six in a hundred answers with a mistake is an acceptable rate for brainstorming and a worrying one for filing a regulatory document.
The release also reflects the commercial logic of the autumn. After Gemini 3 Pro won attention in November, OpenAI faced pressure to show it still led on the professional tasks that drive paid seats. Splitting the launch into Instant, Thinking and Pro variants lets it price those tiers separately, from the $1.75 mainstream rate up to Pro's $21 per million input tokens.
What to watch
- Independent ARC Prize verification of the 52.9% result and its per-task cost.
- Whether developers accept a price increase when rivals are cutting.
- How GDPval-style benchmarks, built by the vendor, hold up against third-party enterprise evaluations.
