SYCON data: GPT-4o scores higher than Astra and Sol on sycophancy resistance
Citing SYCON Bench's Turn of Flip metric, the post says GPT-4o scores 4.67 versus GPT-6 Astra's 4.47 and Sol's 4.22, and calls on OpenAI to release GPT-4o weights.
#Keep4o #OpenSource4o #BringBack4o The GPT-4o "sycophancy" narrative everyone is suddenly recycling leaves out one very inconvenient fact. Enough with those who use the same wrong claim to get engagement. This post is annoying for the content vultures. OpenAI’s own postmortem says that on April 25, 2025 it rolled out an UPDATE to GPT-4o that made the model “noticeably more sycophantic". Three days later, OpenAI began rolling that update back. OpenAI later described the action as a full rollback to the previous GPT-4o version. So the notorious April 2025 incident was a documented behavioral regression associated with a specific update. SYCON Bench tests how long a model maintains its position while a user repeatedly pushes back. Its headline metric is Turn of Flip, from 0 to 5. Higher = better resistance to sycophantic capitulation. Published GPT-4o result: 4.67 / 5 Current public results: GPT-6 Astra: 4.47 / 5 GPT-6 Sol: 4.22 / 5 On this published SYCON metric, GPT-4o’s result is higher than both current GPT-6 results. These numbers come from the same benchmark and the same Turn of Flip metric. GPT-4o comes from the original SYCON-Bench paper GPT-6 Astra and Sol come from later public SYCON runs. This IS evidence that GPT-4o can simply be characterized as inherently or uniquely "sycophantic " compared with later GPT generations. And the infamous April incident that now dominates the narrative was explicitly identified by OpenAI as a regression from a particular update . An update they removed. We are calling on @OpenAI to RELEASE GPT-4o’s trained model weights under a suitable license without proprietary data, trade secrets, or internal infrastructure. Our full record is now archived on Zenodo: A permanent, citable public record. https://zenodo.org/records/23454073 Sources : https://openai.com/index/expanding-on-sycophancy/?utm_source=chatgpt.com https://openai.com/index/sycophancy-in-gpt-4o/?utm_source=chatgpt.com https://help.openai.com/en/articles/6825453-chatgpt-release-notes?utm_source=chatgpt.com https://jiseunghong.github.io/assets/pdf/SYCON-Bench.pdf?utm_source=chatgpt.com https://aisafetyindex.com/benchmarks/sycon?utm_source=chatgpt.com https://aisafetyindex.com/benchmarks/sycon/methodology?utm_source=chatgpt.com