The most prestigious law school admissions discussion board in the world.
|
 |
 
|
Opus 4.5 is disgustingly good
Poast new message in this thread
Date: November 29th, 2025 12:55 PM Author: Lavender Institution Knife Pretty sure humans are effectively useless now other than to pilot AI systems (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470203) |
Date: November 29th, 2025 1:37 PM Author: tantric chapel The level of improvement from Sonnet 4.5 is insane (a model that is a couple months old). LLM progress isn’t really slowing down in a meaningful way. (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470302) |
 |
Date: November 29th, 2025 2:10 PM Author: Lavender Institution Knife yeah and whats also interesting is that sonnet 4.5 was legit good..like it was better than opus 4.1 for most things and competitive with gpt5. I still like gpt 5.1 and the 5 pro api best as a generalist but opus 4.5 is insane for coding and also its much better at critiquing something you give it. the other LLMs are still too nice, but it has no problem giving harsh criticism and arguing against you. GPT 5.1 is better than 5 for this but still a bit weak compared to claude (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470392) |
|
Date: November 29th, 2025 2:12 PM Author: harsh talking lodge (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470405) |
|
Date: November 29th, 2025 2:13 PM Author: Aromatic National Newt My retarded fucking company refuses to upgrade our GHCP models from gpt4o and I want to kill myself (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470410) |
|
Date: November 29th, 2025 2:49 PM Author: Lavender Institution Knife lol jfc (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470540) |
|
Date: November 29th, 2025 2:50 PM Author: bearded talented station (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49470543) |
|
Date: December 7th, 2025 10:17 AM Author: Indecent motley parlour wow (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491040) |
|
Date: December 7th, 2025 10:32 AM Author: bespoke vigorous cruise ship lmao (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491074) |
|
Date: November 30th, 2025 1:14 AM Author: bearded talented station (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49471468) |
|
Date: December 7th, 2025 10:19 AM Author: Indecent motley parlour it truly is insane - what a gift (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491048) |
|
Date: December 7th, 2025 10:23 AM Author: Misunderstood Well-lubricated Sanctuary Selfie explain what cause you to reach this conclusion (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491063) |
|
Date: December 7th, 2025 10:50 AM Author: Maize Principal's Office Mental illness (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491106) |
|
Date: December 7th, 2025 10:44 AM Author: beady-eyed filthpig I need an AI agent to work with (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49491097) |
|
Date: December 21st, 2025 12:59 PM Author: tantric chapel Opus 4.5 is up to 4 hours and 49 minutes on the METR time horizon task. this benchmark measures the task length (in terms of human work time) that models can do with SWE/AI research type projects. big increase over 5.1 max, which was 2 hours 53 minutes and faster than the overall trend of doubling every 7 months. with any luck, models will be capable of substantially automating AI research before 2030 and set off an intelligence explosion.
https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ |
|
Date: December 21st, 2025 1:47 PM Author: exciting university People complaining that 4.5 was shadow throttled starting a few days ago (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49527733) |
|
Date: December 21st, 2025 1:53 PM Author: harsh talking lodge They are always claiming this for new model releases and no objective evidence ever materializes for it. It’s almost certainly a psychological bias rather than reality. There is substantial variation in how models respond to a particular problem just based on how they are prompted. It’s hard for a user to reliably measure model capabilities over time based on intuition alone. (http://www.autoadmit.com/thread.php?thread_id=5804093&forum_id=2/en-en#49527743) |