GPT-5.2 Thinking evals

10 easter eggs to celebrate 10 years. good luck. https://t.co/BPT9kKuenb

10 years.

GPT-5.2 is part of an ongoing set of model improvements. We’re continuing to iterate and work on known issues like over-refusals and latency to make our models more useful. GPT-5.1 in ChatGPT will remain available to paid users for three months as a legacy model.

GPT-5.2 Pro Our smartest and most trustworthy model for difficult questions: - Stronger performance in complex domains like programming - Best model for assisting and accelerating scientists

GPT-5.2 Instant Built for everyday work and learning: - As warm and conversational as GPT-5.1 - Clearer explanations that surface key information upfront - Improved how-tos and walk-throughs - Stronger technical writing and translation - Better support for studying and

GPT-5.2 Thinking Raises the bar for professional work: - State-of-the-art long-context reasoning - Major improvements in spreadsheet creation, analysis, and formatting - Early gains in slideshow creation

On GDPval, an eval measuring well-specified knowledge work tasks across 44 occupations, GPT-5.2 Thinking is our first model that performs at a human expert level. These tasks include making presentations, spreadsheets, and other artifacts.

GPT-5.2 Instant, Thinking, and Pro are rolling out today, starting with Plus, Pro, Business, and Enterprise plans. Free and Go users will get access tomorrow. GPT-5.2 is also now in the API and Codex.

GPT-5.2 is now rolling out to everyone. https://t.co/nfubPwnIIw

As our models grow more capable in cybersecurity, we’re investing in strengthening safeguards and working with global experts as we prepare for upcoming models to reach ‘High’ capability under our Preparedness Framework. This is a long-term investment in giving defenders an

Denise Dresser is joining OpenAI as Chief Revenue Officer. Previously CEO of Slack, she brings deep enterprise and customer experience as she leads our global revenue strategy and support for customers at scale. https://t.co/yQcgkhCdo0

we released our first State of Enterprise AI report today -- grounded in actual usage and survey data, it shows that enterprise AI adoption is not only broadening but deepening. some notes: *enterprise messaging volume is up ~8× YoY, with the avg employee sending ~30% more

I'm seeing lots of confusion about ads rumors in ChatGPT. There are no live tests for ads – any screenshots you’ve seen are either not real or not ads. If we do pursue ads, we’ll take a thoughtful approach. People trust ChatGPT and anything we do will be designed to respect that.

One year ago today, OpenAI launched ChatGPT Pro. I’ve spent the year treating it as my “first hire” at Sound for Movement and keeping field notes on how I use it daily. Grateful to @KevinDubouis and @kevinweil at @OpenAI, and everyone who’s been sharing already:

Confessions don’t prevent mistakes; they make them visible. Next, we’re scaling the approach and combining it with other alignment layers—like chain-of-thought monitoring, instruction hierarchy, and deliberative methods—to improve transparency and predictability as capabilities

In our tests, we found that the confessions method significantly improves the visibility of model misbehavior. Averaging across our evaluations designed to induce misbehavior, the probability of “false negatives” (i.e., the model not complying with instructions and then not

We trained a variant of GPT-5 Thinking to produce two outputs: (1) the main answer you see. (2) a confession focused only on honesty about compliance. The main answer is judged across many dimensions—like correctness, helpfulness, safety, style. The confession is judged and

AI systems are becoming more capable, and we want to understand them as deeply as possible—including how and why they arrive at an answer. Sometimes a model takes a shortcut or optimizes for the wrong objective, but its final output still looks correct. If we can surface when

In a new proof-of-concept study, we’ve trained a GPT-5 Thinking variant to admit whether the model followed instructions. This “confessions” method surfaces hidden failures—guessing, shortcuts, rule-breaking—even when the final answer looks correct. https://t.co/4vgG9wS3SE