Bijan Bowen says Claude Haiku 5.5 has pretty insane capability; tests used 1% of weekly usage

The OS had no search option, and Bowen called its wallpapers ugly. A map overlay obscured part of one game, though the driving and paint apps worked well enough for Bowen to consider the OS acceptable for a model in this class.

A second game in the OS was unexpectedly 3D. Bowen pointed out an effect when the ship was hit and called the game simple but not bad.

The revised skate game came close to six-month-old state of the art

The first C++ skate game, built without dependencies, took just under 38 minutes. It used arrow keys instead of the movement keys Bowen expected, included a replay feature and showed the skater’s trajectory. The board moved with the player during tricks. Bowen called the game basic but said it was better than GPT-6 Luna’s skate-game result.

Bowen then gave the model feedback and a reference image from the Claude Sonnet 5.5 skate-game review. The overhaul took 23 and a half minutes. It brought in the boardwalk and pedestrian movement shown in the reference, but the skater’s legs still moved with the board during tricks, and some storefront text appeared inverted.

Bowen nevertheless called the revision a remarkably strong result for a Haiku model, close to state of the art from about six months earlier. They praised its buildings, movement and boardwalk, along with the skybox, traffic, wipeout replay and water splash. Bowen also compared it favorably with Claude Sonnet 5.5’s skate-game result. “Overall, I'll say this is quite quite impressive, especially considering this class of model”, Bowen said.

The pool-party game was one of the best showcases

Bowen described the backyard pool-party game as a relatively new task, less likely to have appeared in training data. The game was made with Blender and Godot. Bowen initially did not know how to play, then found the practice mode and pumping controls.

The water effects, a dog jumping into the pool, string lights and options screen stood out to Bowen. They also mentioned some background coloring, but called the game one of Haiku 5.5’s best showcases and a good result for a model in its class.

Feedback made the subway game worth the extra time

The first subway first-person-shooter game, made with Claude Code, took 35 minutes. Bowen noted a weapon-rendering problem and freezing when the train arrived, while saying the shooting logic worked and the result was not bad for a Haiku model.

A revision took eight and a half minutes. It fixed the floating-hand issue, but the game still froze when the train was due. Ammunition tracers remained on screen; Bowen thought they looked cool, while also treating their persistence as an issue. After further feedback about the freezing, Bowen said it appeared to have stopped.

Bowen praised the revised game’s stations and their different aesthetics, the information screen inside the train and the result shown when the player lost. They called the version clean and well done, and said, “This was definitely worth spending a little bit of extra time on.”

Bowen found the watch quite nice and the derby interesting

For a RuneScape 2007 test, Bowen used Ultra Code to request a pixel-perfect replica of PvP at the Grand Exchange. The result showed a bronze sword and beginner-level character stats; Bowen said they did not see NPCs.

The watch-website test took 22 minutes and nine seconds at maximum effort. Bowen requested an image as the watch dial and an exploded view. They liked the details and exploded view, but noted narrow bands, a blurry-looking special-edition face and that the watch could not be manipulated manually. Bowen still found the result quite nice.

The final test used an image generated with Gemini as a reference for a demolition-derby game made with Blender and Godot. Bowen requested soft-body physics, camera options, falling parts and suspension. They called the cars hilarious-looking, but found the arena and suspension decent. Collisions produced smoke particles, and Bowen thought a bumper appeared to bend, though the soft-body physics were hard to discern. They called the result interesting.

Bowen closed by returning to the usage figure: “the entirety of this testing used 1% of our weekly usage.” They called that “That's insane, I think” and ended on Haiku 5.5’s pretty insane capability.