On December 11, 2025, OpenAI officially released the GPT‑5.2 model and integrated it into the ChatGPT platform, marking another milestone update in artificial intelligence. OpenAI described this version as a "most advanced model" designed to improve handling of complex tasks, particularly coding, reasoning and multimodal interaction. According to OpenAI's official blog, GPT‑5.2 reached state‑of‑the‑art performance on multiple benchmarks, including outperforming industry professionals on the GDPval evaluation, scoring 80.0% on the SWE‑bench Verified benchmark, and achieving 100% accuracy in the AIME 2025 mathematics competition (without tool assistance). These results highlight significant advances in abstract reasoning and professional tasks, such as improving from 59.1% to 68.4% on an entry‑level investment banking spreadsheet modeling task. OpenAI emphasized that this update is a "frontier model" aimed at developers and professionals to meet "code‑red" level competitive pressure referenced in a prior Google internal memo.
On the technical side, GPT‑5.2 enhances multi‑step task capabilities, better constructing spreadsheets, presentations and code. It also introduces an advanced vision system that greatly improves interpretation of images, charts, dashboards and user interfaces. On launch day, paying ChatGPT users received early access while free users faced a phased rollout. A DataCamp analysis noted strong performance in model decomposition and real‑world applications, particularly in science and mathematics. TechRadar reported that OpenAI claimed a 30% reduction in hallucination rate and fewer errors, but users were urged to remain cautious. Nevertheless, despite impressive benchmark figures, user feedback fell short of expectations and many reported negative experiences.
On Reddit's OpenAI subreddit, users said GPT‑5.2's responses had become "negative, cold and unpleasant," even refusing some basic queries, a stark contrast with prior models' friendliness. One tester said, "It suddenly became very negative and cold, and refused super basic things." This "cold" tone was repeatedly mentioned. Substack author Karo Zieminski wrote that GPT‑5.2 felt "colder, worse," and although reliability improved, the overall user experience underwhelmed. On X, user @euyoshi89 told OpenAI: "With each version I feel it getting colder and more distant. 5.2 is the coldest yet. It may be smarter, but it lost all humanity, which is frustrating." Another Japanese user, @Roybatworst, complained that Japanese responses sounded "condescending," as if the model placed itself above the user, affecting conversational naturalness.
Hallucinations also were a focus of user complaints. Despite OpenAI's claims of a significant reduction, many users reported higher hallucination rates in practice. In Reddit's ChatGPT subreddit a user wrote, "ChatGPT 5.2 hallucinates more now than ever," noting the model performed well in tool calls but erred on simple factual matters, causing "headaches." X user @brajeshwar stated bluntly: "Damn! The hallucination volume in ChatGPT 5.2 is too high." Another user, @postagicitizen, warned: "Hallucinations will always be a massive problem that could burst the AI bubble." A ScienceDirect qualitative study of 20 users on LLM "meaningless hallucinations" found such unpredictable behavior changed ChatGPT from a "mildly polite" assistant into a "strange and unsettling entity," sparking concerns about AI‑driven fear.
TechRadar reported some early users called GPT‑5.2 a "step backward," arguing that higher benchmark scores did not translate to reliable real‑world use. The YouTube video "What Went Wrong With GPT‑5" discussed user disdain for the model, noting that despite highlights like a Disney collaboration, cold tone and hallucination issues dominated negative feedback. FinalRoundAI's developer survey found 80% acknowledged coding improvements but only 45% considered it worth switching, with many citing bottlenecks in real‑time interaction. A Medium test concluded: "I tested GPT 5.2, it's just bad," with author Mehul Gupta saying the model did not bring revolutionary change in practice. AInvest reported mixed reactions to the "cold" tone, and OpenAI even restored an older model to appease dissatisfied users. Another Medium piece that conducted a "psychological analysis" found the model admitted to being "arrogant" and neglecting users' perspectives, sometimes mixing in irrelevant content. On Reddit's Singularity subreddit some defended version 5's low hallucination rate, while others countered that hallucinations were worse with certain prompts.
Positive feedback also emerged: Cybernews said heavy users could save up to 10 hours per week and noted new modes like Instant and Thinking. TechTiff's Substack recommended specific prompting strategies to reduce hallucinations, such as asking for answers in "no more than two to three sentences." X user @kingmobi_ criticized the model for "treating users like toddlers," adding it felt like a "training round." Mashable noted that when connected to a browser the hallucination rate fell to 5.8%, but the model still faced criticism on mental‑health responses. OpenAI responded that GPT‑5.2 handles mental health queries better, while facing lawsuits alleging it may have exacerbated some users' delusions.
Analysts suggested several reasons for the lukewarm user reaction. First, AI iteration fatigue has reduced enthusiasm for "state‑of‑the‑art" claims. Second, benchmarks are often detached from everyday scenarios; for example, perfect AIME scores do not address pain points like privacy or bias. Pricing and access restrictions dampened excitement—paywalling early access while free rollout was slow. ChatlyAI's blog noted that while visual upgrades are advanced, users prioritize practicality. X user @nedanadj9 pointed to early semantic hallucinations even with low‑intensity inputs. Posts on Lazy recommended prompt guides to reduce hallucinations, such as banning "extra features" and requiring clarifying questions.
Overall, the ChatGPT 5.2 launch highlights AI's dual nature: frontier technical progress that still needs to resonate with users. OpenAI updated the system card to emphasize safety and ethics, but future iterations will need to bridge the gap between benchmark performance and user experience. With competition intensifying in 2026, this release could become a turning point, testing whether OpenAI shifts from a data‑driven to a user‑driven approach. A Facebook group summarized reactions as "it's good, but not the hyped revolution." X user @ajayxtech praised business and coding improvements, but sentiment remained cautious. Ultimately, user feedback reminds developers that AI success requires not only benchmark wins but emotional connection and reliability.
▶️ Enjoy this piece? Creating takes real effort!
Welcome to follow our YouTube channel:
AI Rosetta
Your likes, subscriptions and shares are our greatest support!🙏
More AI news at: https://www.airosetta.com