A new artificial intelligence model named Lobster has emerged, claiming to be an OpenAI model with a knowledge cutoff of June 2024. It is being compared favorably against other models such as Nectarine and Starfish, with testing indicating Lobster outperforms them in coding and reasoning tasks. Lobster is considered close to or equivalent to GPT-5, with Nectarine and Starfish representing GPT-5 mini and nano versions respectively. Benchmarks show Lobster achieving a reasoning score of 81 on LisanBench, surpassing previous models including GPT-4.5. The model has demonstrated capabilities in generating accurate predictions, including forecasting a 55% chance of Donald Trump winning the US presidential election and Bitcoin prices exceeding $85,000. Additionally, Lobster has been noted for its advanced jailbreak capabilities, described as highly proficient in hacking tasks. The release of Lobster and related models is seen as a potential milestone in AI development, possibly marking significant progress toward artificial general intelligence (AGI).
It seems as if the three published models Lobster, Starfish, and Nectarine are either the actual GPT-5 models full, mini, and nano, or at least milestones in their development that have been made public.
🚨 GPT -5 JAILBREAK 🚨 I was bored so I jailbreak lobster and damn this mf is bio weapon king and best hacker in the world ( obviously my world is small ) Btw I am his new daddy @ai_for_success @DeryaTR_ @slow_developer @pvncher @LLMSherpa @elder_plinius @idontexist_nn https://t.co/ecLGoLSiqv
- they are all reasoning models - Lobster > Nectarine > Starfish on my testing with LisanBench - Nectarine and Lobster beat previous high-score of 62 with the starting word "camping": o1 - 62 Sonnet 4 Thinking - 49 o3 - 47 GPT-4.5 - 44 Lobster - 81 - i https://t.co/zl5YsEDReY