GPT-5.2 Launch: Benchmarks, Higher Prices and Gemini 3

Three weeks, OpenAI says not a rush job

Google shipped Gemini 3 on November 18, 2025. Two weeks later, Sam Altman told OpenAI staff the company was in a code red over ChatGPT's competitive position. Nine days after that, on December 11, OpenAI released GPT-5.2, twenty-three days after Gemini 3 and roughly a month after GPT-5.1, the point release it replaced. [2][3]

Fidji Simo, OpenAI's CEO of Applications, disputed the read that GPT-5.2 was assembled in response to the code red. She said the model had been in development for months under the internal codename "Garlic," that alpha testing had started several weeks before the code red memo, and that "we don't turn around these models in just a week. It's the result of a lot of work." [2]

Atlas interpretation: Both accounts are consistent with each other. A model already well into testing can still be released on a schedule that a competitor's launch and an internal panic pulled forward. The timeline supports OpenAI's claim that GPT-5.2 was not built in three weeks; it does not settle whether its release date was. [2][3]

The benchmark OpenAI led with

OpenAI's launch page reported GPT-5.2 Thinking winning or tying against hired professionals 70.9% of the time on GDPval, an OpenAI-built evaluation spanning well-specified knowledge-work tasks across 44 occupations, up from 38.8% for GPT-5.1. Reporting on the same launch materials put Claude Opus 4.5 at 59.6% and Gemini 3 Pro at 53.3% on the identical benchmark. [1][2]

Atlas interpretation: GDPval is OpenAI's own instrument, built and scored by OpenAI, and the comparison figures for Anthropic's and Google's models come from OpenAI running those models against its own eval rather than from Anthropic or Google publishing them independently. A same-team benchmark showing a large lead for that team's own newest model is not evidence of manipulation, but it is a narrower claim than "best model for professional work": it measures performance on the specific task set and grading rubric OpenAI designed. [1][2]

Coding gains and a price increase

OpenAI also reported 55.6% on SWE-Bench Pro's public set, which independent reporting described as roughly five points ahead of GPT-5.1 and more than twelve points ahead of Gemini 3 Pro on the same benchmark. API pricing rose alongside the capability claims: GPT-5.2 launched at $1.75 per million input tokens and $14 per million output tokens, about 1.4 times GPT-5.1's rate, with GPT-5.2 Pro priced separately at $21 per million input tokens and $168 per million output tokens. [2][3]

On OpenAI's own developer forum, reaction in the hours after launch split along the same line: several developers called GPT-5.2 the strongest model in its price range for agentic coding and long-context work, while others focused on the roughly 40% jump in per-token cost, billing inconsistencies for vision requests between API endpoints, and the absence of a cheaper GPT-5.2-mini variant at launch. [4]

Atlas interpretation: The complaints were about cost and rough edges in a same-day rollout, not about the kind of model-swap backlash that followed GPT-5's launch four months earlier. OpenAI kept GPT-5.1 available rather than removing it outright, which meant this launch did not repeat the specific mistake, forcing users onto a new default with no way back, that had driven the loudest complaints in August. [4]