In a video review by Bijan Bowen, Bowen ends up fairly disappointed with Step 5 Preview. Recurring spatial and geometric problems undercut several projects, and Bowen saw no reason to choose the model over DeepSeek’s similarly sized Flash model.
The tests ranged from a generated desktop and several games to a watch website. Some details drew praise, but the problems accumulated: Bowen also rated GLM’s Flash model better in the tests and said the earlier Step 3.7 Flash had impressed them more.
The browser OS was acceptable, despite its glitches
Bowen began with a generated desktop environment, moving through its apps, wallpapers, menus and file browser before launching a game. The file browser felt full-featured, and the interface drew a clear compliment: “This is a good looking UI, actually”.
But the game appeared inverted, and Bowen found moving its windows frustrating. The overall result was acceptable, they said, with some promising touches alongside those glitches.
The skate scene had more flaws than successes
For a C++ skate game set on a New York City block, Bowen specified that Raylib could not be used. The skaters were assembled from separate pieces, and the scene included a newsstand and a fire hydrant spraying water. Bowen also pointed to objects they thought might be vehicles and a pedestrian.
The separate skater parts reminded Bowen of an approach they had seen in a Claude Opus 5.5 result. But the test had more problems than successes, in their view; they described the scene as resembling a collection of assets laid out for inspection. The individual elements did not make the skate game a convincing whole.
The Subway FPS stayed frustrating after an update
In the newer Subway FPS test, players were supposed to board a train to reach later waves. During the first run, the player character was stuck below the floor. Bowen still liked the enemies, which had more detail than capsule-like figures and looked good for a browser-based 3D game.
An update improved the scene, adding glowing ammunition crates and a rat that Bowen singled out as a good detail. But the train-boarding control did not work, so the test stopped there. Bowen said the initial attempt had taken more than an hour without reaching a decent result: “this took over an hour and just did not get us to a decent point”. The remaining problems left them frustrated.
Geometry kept the derby and watch from coming together
Bowen asked for a playable recreation of an image, with sedans modeled in Blender and the game built in Godot. The result included different camera views, but the attempts did not come together. Bowen said that run cost $6.36 and made them reluctant to spend more through OpenRouter.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
A watch website offered more individual details Bowen liked, including parts of the case and strap, the front end and the watch face. It also had an exploded view and a special edition. The exploded-view presets appeared as working options, but the geometry was often crooked or misaligned, including the sapphire glass. Bowen treated those flaws as a major letdown for the site overall.
The RuneScape replica had game-specific details but failed
The final game request was for a playable, pixel-perfect replica of RuneScape 2007 PvP at the Grand Exchange. Bowen was surprised that the result used Godot. It included relevant game lore and some combat logic, and Bowen judged it reasonably well made as a small 2D game.
Still, the game lacked an inventory and a proper user interface, and it kept glitching and crashing. Bowen did not consider it a successful result.
Bowen preferred other models
Across the projects, Bowen’s main complaint was a recurring lack of spatial and geometric coherence. “This model seriously has some like deep geometric or spatial issues and that's very much let these results down,” they said.
Bowen put the total cost of the tests at a little over $22 and considered that expensive for the results. They cited prices of $1 per million input tokens and $2.70 per million output tokens, which they said were increasingly hard to compete with when models such as GPT-6 Luna and Claude Haiku 5.5 were available at similar prices.
The comparison was part of Bowen’s overall verdict: they preferred the earlier Step 3.7 Flash, saw GLM’s Flash model as a better option based on these tests, and did not see a reason to choose Step 5 Preview over DeepSeek’s similarly sized Flash model. Bowen allowed that Step 5 Preview might have niche uses and said its planned open-weights release on Oct. 15 could be useful. They hoped it would use the Apache 2.0 license, as they said Step 3.7 Flash did, but noted the license was not yet known.
Bowen’s closing judgment was: “I have to say overall I am fairly disappointed”.
Comments
Comments are written by readers. They are not reporting or opinion from AI News.
Share your view on this story. Criticise ideas and public records, not other readers.
Comments appear once a moderator approves them.
Community guidelines
More in our terms and privacy policy.
Comments close on October 19, 2026.
No comments yet. Start the conversation.