In Bijan Bowen’s YouTube video review of Claude Sonnet 5.5, the verdict is emphatic: Bowen calls the model a monster and says it absolutely demolishes GPT-6 Sol on the visual tasks they tested. A GTA-style game outdid a result from the pricier Claude Opus 5.5, while an unfamiliar pool-party game answered Bowen’s question about whether Sonnet 5.5 could excel beyond familiar tests.
Bowen described Sonnet 5.5 and GPT-6 Sol as direct competitors at the same price, with Opus 5.5 costing twice as much. That price gap mattered to their assessment of the games. The review also put the model through a robot-arm task and a one-sentence request for a RuneScape 2007 replica.
A long run produces a GTA-style game Bowen calls too good
The browser-based operating-system run continued for more than an hour, then reached about two hours without finishing. Frustrated, Bowen asked the model for the file so they could inspect what it had made. They dropped another planned game after the wait and said they had learned not to run the model at maximum effort.
Bowen found the initial background hideous and repulsive. A different seed produced a background they found more appealing. “That's too good,” Bowen said of the result, judging the GTA-style game made within the browser OS better than Claude Opus 5.5’s version.
The game had working traffic lights, an emergency brake and multiple camera views. Bowen drove through a setting with a taxi and a delivery van, then explored a dock area with shipping containers. They pointed out one missing detail: bullet holes.
The comparison with Opus was direct. Bowen found Sonnet 5.5’s GTA-style game better and questioned whether it was worth paying twice as much for Opus on the kinds of tasks they had tried. For this run, they said, the result was worth the frustration.
A skate city bigger than any Bowen had seen
The C++ skateboarding game produced a city map Bowen said was bigger than any other result they had seen in these tests. It had a boardwalk, buildings, moving traffic, boats, collectibles and a minimap. Pedestrians commented on the skating as the player moved through the city.
During troubleshooting, the model switched to high effort. Bowen found the cars odd-looking, though not bad, and noted that the game lacked sound. Their attention kept returning to the size of the map and how much was happening across it.
Bowen said the skate-game result might be at least as good as Claude Opus 5.5’s, or better. They preferred Opus’s water and possibly its distant buildings. Even with that comparison, they called Sonnet 5.5’s map really well done and said it exceeded their expectations.
A chaotic robot-arm record and an exceptionally well-done subway game
The robot arm reached its goal by an unlikely route. It moved erratically, pushed against the table and lifted its own base before grasping and moving the car. Bowen called the outcome “task failed successfully.”
Sonnet 5.5 beat the previous quickest result in Bowen’s test by more than 20 minutes. Bowen put that earlier time, shared by GPT-6 Astra and a Gemini Flash model, at about 40 minutes. They said the arm had no business completing the task the way it did.
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 viewer suggested a variation for the subway first-person shooter: after clearing one station, the player would board a train and reach another station with enemies. Bowen ran it at maximum effort for a little under two hours. The game included smoke, bullet holes, health packs and the train sequence the prompt requested.
Bowen initially mistook sounds from another project for the subway game’s audio, then corrected that impression. They also found that the train handles stayed still while the train moved and that the lights could not be shot out, unlike in a GPT-6 Astra test they recalled. “That was exceptionally exceptionally well done,” Bowen said of the subway result.
A new pool-party idea answers Bowen’s question
Bowen wanted to know whether Sonnet 5.5 was particularly good at tasks they had tried before or could handle an idea new to their tests. They devised a Backyard Pool Party prompt during the review and switched the run to Ultra Code.
An early playable build included a backyard pool, diving board, selectable players, a dog and a barbecue. Bowen singled out the water effects and laughed at some poorly executed dives. They called the result absolutely bonkers and very well done. For Bowen, this unfamiliar game answered the question: Sonnet 5.5 was genuinely good.
Bowen also used Ultra Code for a one-sentence prompt requesting a pixel-perfect replica of RuneScape 2007. They tried movement, spells, player-versus-player combat and loot. Arrow keys worked, while W-A-S-D movement did not; a special attack was unavailable at the displayed 30% level.
The combat effects impressed Bowen. They said Sonnet 5.5 had come closer to a pixel-perfect RuneScape 2007 replica than any model they had tested on that task before.
Bowen read figures from the model’s introduction post describing Sonnet 5.5 as 30% faster and up to 30% lower in cost for more work than Sonnet 5. They said the review had used less than 10% of their weekly usage, though they could not recall the starting figure.
Bowen’s closing verdict was that Claude Sonnet 5.5 had exceeded their expectations to an incredibly high degree. They called it a fantastic model and said, “this model is absolutely a monster.”