GPT-6.1 Sol Is So Weird

My Patreon - https://www.patreon.com/cw/CanItCode Two new models in two days: Claude Sonnet 5.5 from Anthropic and GPT-6.1 Sol from OpenAI. Sonnet is ahead of Opus on Terminal-Bench, and both new models cost $2 per million input tokens and $10 per million output tokens, half the price of Claude Opus 5.5. So I put all three on the same dungeon crawler, in three rounds: first a low-poly dungeon kit in Blender, then a knight and three monsters (a skeleton warrior, an acid slime and a giant cave bat), modelled, rigged and animated from scratch, and finally a small playable game in Godot built only from their own work. No Blender MCP: the models write Python scripts, run Blender headless and work in a gauntlet loop, where one agent builds and one critic checks the results. The model decides on its own when it is done. Round one: Opus 5.5 came closest to the reference, Sonnet 5.5 right behind it, GPT-6.1 Sol flatter, but it finished in 80 minutes for under $4 while the others took almost five hours. In the game, Opus 5.5 is simply the most pleasant to play, Sonnet 5.5 is the big surprise of the episode, and GPT-6.1 Sol got worse every round (its special attack is pure Houdini). Along the way: why one hit feels good and another doesn't, with every feedback layer of one sword hit switched off and added back one at a time. The costs are estimated API costs at standard token prices, measured from the session logs (the main agent plus its helpers): Sonnet 5.5 cost $318.00 for all three rounds (13 h 46 min of work), Opus 5.5 $376.30 (12 h 58 min) and GPT-6.1 Sol $29.17 (7 h 5 min). Sonnet is half the price per token, but it used so many more tokens that its bill ended up almost the same as Opus. Source files for patrons (all three games, the Blender kits and characters, and the briefs and prompts the models got): https://www.patreon.com/c/canitcode CHAPTERS 00:00 Two new models in two days 00:09 Benchmarks and API prices 00:33 Three models, three rounds 00:48 Source files on Patreon 01:08 The reference: a flooded prison 01:21 Round 1: a dungeon kit in Blender 02:13 Round 1: cost and time 02:34 Why Sol cost so little 03:01 Inside the wall 03:44 How they did it 04:12 Round 2: a knight and three monsters 05:11 Round 3: the games 06:06 Why a hit feels good 06:28 Sum up and the final bills CREDITS Benchmarks: Artificial Analysis, https://artificialanalysis.ai (screen recording). Launch pages: Anthropic and OpenAI (screen recordings). Sound effects used in the games: Pixabay, Pixabay Content License. Godot, Blender, Pixabay, Claude, Anthropic and OpenAI names and logos are trademarks of their respective owners. Music: "Snowline Drift", made with Suno.