Bijan Bowen evaluates GPT-6 Luna through a browser desktop, several gamesGame generation uses AI to create or assemble playable assets, scenes, rules, audio and interactions., a watch website and a physical robot-arm task, using maximum reasoning effortReasoning effort is the amount of internal computational work an AI model applies before producing an answer or action. throughout and adding fast mode for the robot test. He presents the model as a cheaper successor to GPT-5.6 Luna with modest benchmark improvements, and reports API prices of 10 cents per million input tokens and 50 cents per million output tokensToken pricing is the rate an AI provider charges for processing input tokens, generating output tokens, or reading cached tokens.. Those prices and benchmark comparisons are part of his account rather than an independently verified cost or capability assessment.
Bijan Bowen reports an initially incomplete browser desktop, followed by a repair that restores its animated wallpaper while leaving some interface problems. The C++ skate game is his strongest result: it compiles when he runs the supplied command, and a follow-up adds a more cohesive rainy, neon setting. A subway shooter also improves after a detailed revision requestIterative refinement improves an AI result through repeated cycles of building, inspection, feedback, and revision., gaining clearer visibility, more developed enemies and bullet-hole effects. A separate C++ rally game offers a usable starting point but has nonworking gauges and awkward track rendering, showing that the stronger skate result does not carry evenly across tasksModel reliability is the degree to which an AI model produces dependable behavior across repeated, varied and operationally realistic use..
Bijan Bowen finds the wrestling game playable but visually simple, with limited character movement and no impact sound effects. The watch website includes a working exploded view and an appealing personalized design, but some geometry hides its watch hands. The robot-arm test fails to move the toy car successfully. He describes an unusual camera dashboard and difficulty interpreting depth from a fixed camera angle, then stops the attempt when the arm begins moving in ways he considers potentially damaging.
Bijan Bowen concludes that GPT-6 Luna offers decent foundations that need refinement, especially for repetitive work that demands less sophisticated reasoning. His weekly allowance display remains at 100 percent throughout the tests, which he treats as a sign of efficiency; an unchanged displayed percentage does not establish zero consumption or a measured per-task cost. His recommendation rests on the reported price and acceptable prototype results, while acknowledging that the aesthetics trail the more capable models he has recently tested.
Watch on YouTube




