OpenAI launches GPT‑5.2 with lower hallucination rates and higher

OpenAI announced the immediate rollout of its GPT‑5.2 model series, a set of upgrades to the existing GPT‑5 lineup that the company touts as its most capable for professional knowledge work [1]. The launch, which begins with paid ChatGPT subscribers, includes three variants—Instant, Thinking, and Pro—and claims notable gains in accuracy, safety, and benchmark performance [2].

Key takeaways

• GPT‑5.2 is released in three versions: Instant, Thinking, and Pro [1]
• Hallucination rates drop to 10.9 % for the Thinking variant, versus 16.8 % for GPT‑5.1 [1]
• Benchmark scores improve across coding, math, and science tests (e.g., AIME 2025 math 100 % vs. 94 % for GPT‑5) [2]
• Access starts with Plus, Pro, Go, Business, and Enterprise plans and will expand gradually [2]
• OpenAI says the model is safer for users showing signs of mental‑health distress [1]

New model series and rollout plan

OpenAI’s blog post frames GPT‑5.2 as a “series of models” built on the broader GPT‑5 architecture, positioning it as the most capable iteration for professional tasks [1]. The company will deploy the new series first to paid users, with a phased rollout that may leave some subscribers waiting for availability [2]. Existing GPT‑5.1 models will remain accessible for three months before being sunsetted, allowing a transition period for users [1].

Performance gains and safety claims

According to OpenAI’s system card, GPT‑5.2 shows across‑the‑board improvements on internal and public benchmarks. Notable results include a 70.9 % score on the GDPval benchmark (up from 38.8 % for GPT‑5) and a perfect 100 % on the AIME 2025 math test [2]. The Thinking variant’s hallucination rate is reported at 10.9 %, a reduction from the 16.8 % seen in GPT‑5.1 Thinking, and further drops to 5.8 % when the model accesses the web [1].

OpenAI also highlights enhanced safety for conversations involving mental‑health concerns, claiming fewer undesirable responses in both Instant and Thinking variants compared to earlier models [1]. Mashable notes that these claims have not been independently verified and that the system card provides limited detail on the safety testing methodology [1].

Why it matters

The GPT‑5.2 launch underscores OpenAI’s effort to stay ahead in a competitive AI landscape that includes rivals such as Google’s Gemini series [1]. By emphasizing reduced hallucinations and better benchmark performance, OpenAI aims to reinforce confidence among professional users who rely on accurate, reliable outputs. The company’s focus on mental‑health safety also reflects growing scrutiny over AI’s role in sensitive user interactions. As the rollout proceeds, user feedback and independent evaluations will be crucial to determine whether GPT‑5.2 delivers on its promised improvements.