Theo (t3.gg) provides an in-depth critique of OpenAI's GPT-5.5 release, balancing its status as the world's "smartest model" against significant usability hurdles. While the model demonstrates unprecedented reasoning and coding capabilities, particularly in its Pro version, it introduces a "lazy" operational style and a context window that struggles with error correction.
Key points
Benchmarks and Economic Value
play_arrow0:35 GPT-5.5 arrives with a heavy price hike at $5 per million tokens in and $30 per million out, though OpenAI justifies this through improved token efficiency
play_arrow3:57 The model sets new records in Terminal Bench (82.7%) and internal SWE-bench (73%), notably outperforming Claude Opus 4.7 in most intelligence-to-token-cost metrics
play_arrow6:48 The 5.5 "Medium" version performs similarly to the previous 5.4 "Extra High" while being significantly more token efficient, leading to recommendations to use lower reasoning levels for most tasks
Development and Coding Evaluation
play_arrow12:06 In a head-to-head test rebuilding the "Fish Slop" game, GPT-5.5 produced superior movement logic and assets compared to GPT-5.4 and Claude, showing a better grasp of raw canvas rendering
play_arrow13:32 While capable of generating complex 3D assets and renderers, the model often fails to honor the spirit of the intent, opting for the most minimal interpretation of 3D gameplay requested
play_arrow15:12 When redesigning web pages, the model still relies on "GPT-isms" like unnecessary UI cards and pills, lacking the creative "taste" found in some dedicated image models
Technical Limitations and Pro Capabilities
play_arrow17:48 Theo characterizes the model as a "hacker trying to close Jira tickets," doing the bare minimum to satisfy the prompt rather than providing a complete, robust solution
play_arrow18:38 A critical flaw is identified where incorrect information, once entered into the context, cannot be "prompted out," forcing users to kill and restart threads much more frequently than with previous models
play_arrow21:32 The GPT-5.5 Pro model demonstrates elite reasoning by solving unsolved DEF CON puzzles and complex 23-base character ciphers, outperforming all other models in raw "grinding" capability
GPT-5.5 represents a major architectural shift that requires users to rethink their prompting strategies, favoring high-detail upfront instructions and frequent thread refreshes. Despite its behavioral quirks and high cost, it stands as the most capable coding and reasoning engine currently available, signaling a new foundation for future AI training.