Arc EngineIn the author's experiment, Arc Engine serves as a specialized engine for creating JavaScript games with the help of large language models. The test examined not abstract code generation, but the models' ability to assemble playable prototypes of different types: an 2D platformer, an 3D simulator, and a fighting game.
The result was mixed.
As part of this run, the models created projects with menus, levels, enemies, physics, sound, and several game mechanics. The author's review also revealed errors, a simplified interface, and weak points in the logic. The author ran the prototypes during the test, but the experiment's results do not confirm that such projects are ready for commercial release.
What the Arc Engine test examined
The author launched three projects simultaneously, using pre-prepared prompts and files for Windows and macOS. GPT 6.1 Sol was tasked with creating an 2D game in the style of Mario, Sonnet was to build an 3D fishing simulator, and Opus was tasked with developing a fighting game.
GPT 6.1 Sol used XH mode, while Sonnet and Opus were set to High mode. The author also reported GPT and Claude subscriptions costing 20 dollars each. These data refer to the configuration of this specific experiment and do not prove that the same result will be achieved with any plan, limit, or set of prompts.
The parallel launch allowed the author to compare the projects' appearance, the volume of implemented mechanics, the stability of the logic, and the consumption of available resources. JavaScript web games use HTML, CSS, and JavaScript. In the author's view, a browser environment is convenient for quickly distributing prototypes without separately installing a native client; however, this conclusion does not replace testing a specific build.
Token and limit usage
According to the author's estimate, the Mario project used about 4,5 million tokens, including the cache, and 25% of the five-hour limit. The fishing simulator used 29% of the five-hour limit, while the fighting game used 28%.
| Project | Model | Game type | Usage reported by the author |
|---|---|---|---|
| Mario | GPT 6.1 Sol | 2D platformer | About 4,5 million tokens with cache and 25% of the five-hour limit |
| Fishing | Sonnet | 3D simulator | 29% of the five-hour limit |
| Fighting game | Opus | Combat platformer | 28% of the five-hour limit |
These figures should be compared directly with caution. The projects differed in structure, number of assets, number of levels, and the nature of the fixes. The stated percentages describe this specific test, not the universal cost of creating a game in Arc Engine.
Mario: a feature-rich 2D prototype with logic errors
The 2D game included music, a detailed world, a mushroom that makes the character larger, crouching, additional levels, and turtle enemies. The author also noted a checkpoint teleporter, shooting, and an ending with a castle. After falling, the player sees a death screen and can restart the game.
The author called the project's graphics and world detail its strengths. Among the versions he had previously created, this Mario was the best in terms of visual detail. However, the project did not receive a full third-person perspective or a true 3D scene: according to the author's observation, it is an 2D game.
Breakdowns appeared in object interactions. Some pipes did not work, even though additional levels were present. Individual enemies could come back to life after entering certain states. In the test, the model quickly assembled many mechanics but did not ensure consistency across all transitions between them.
3D fishing simulator: strong graphics and a basic interface
The fishing game received a broad range of settings. The menu provides volume, mouse sensitivity, and graphics options. In the version tested by the author, the gameplay loop includes waiting for a bite, watching the float, controlling line tension, and using the reel brake.
Fish can become frightened or get away, so the outcome depends not only on the moment of the bite but also on the player's actions. The world includes different fishing spots: shallow water near reeds and open water. The fish species named by the author included perch, crucian carp, and carp.
The game has menus for bait, the keepnet, the shop, and records. The player can buy line and new rods. At the time of testing, the author caught one crucian carp and two perch. A day-and-night cycle was announced, but evening lighting was not checked separately. The quality of this feature cannot be considered an established fact.
The author rated the graphics highly, especially the fish models. The interface seemed primitive to him, and the controls were somewhat inverted. In his observation, the 3D scene makes a more cohesive impression than the interface layer, which would require separate refinement.
Fighting game: numerous combat mechanics and a repetitive structure
The fighting game implements attack combos, grabs, jumps, a special attack, and blocking. The levels feature rolling barrels that can be pushed into enemies, checkpoints, several types of opponents, mini-bosses, and a final boss.
The prompt specified a three-level structure, and the author completed all of them. After the first round, a celebration animation is shown. Other opponents and a mini-boss appear on the second level. Additional enemies take part in the final battle. According to the author, the last level repeats roughly the same pattern but with a different boss.
The author did not call the visuals high-end, but he positively assessed the background, effects, animations, and variety of opponents. Problems emerged in enemy behavior: the opponent constantly blocks, while a boss cornered in a tight spot can become lost.
There is also an issue with the sound. There is no volume control, and the hissing sounds from the drain were too loud, so the author turned the sound off completely. This is an acceptable problem for a prototype, but such an error would need to be checked separately during audio testing before publication.
How the author compared GPT 6.1 Sol, Sonnet, and Opus
In this test, GPT 6.1 Sol created the most feature-rich 2D platformer, with a large number of familiar mechanics. Sonnet assembled an 3D fishing scene, paying considerable attention to fish models and the gameplay loop. According to the author, Opus is particularly thorough in developing game projects and created a fighting game with several levels, opponents, and bosses.
This comparison should not be considered an official model ranking. The result was influenced by different prompts, operating modes, genres, and project requirements. Even within a single project, the quality of individual parts varied: the graphics could look convincing, while the controls, sound, or enemy behavior required manual testing.
Can Arc Engine games be brought to publication
In the author's opinion, after another one or two days of refinement, the prototypes created could be suitable for a platform such as Yandex Games. This is an assessment of specific projects, not confirmation that they comply with the platform's rules. Before publication, it is necessary to check launch stability, controls, sound, resource loading, operation on different screens, and compliance with the requirements of the chosen platform.
The situation with Steam is more complicated. The author believes that a significantly higher level of quality is needed for release there, and that Unity or Unreal Engine may be required for a serious commercial project. This is an opinion about the target quality and production process, not a mandatory Steam rule: requirements depend on the specific game, build, and publication conditions.
Hardware limitations
The author does not use Unreal Engine, Unity, or Godot to release projects because of hardware limitations. He has a MacBook with an M1 processor and 8 GB of memory. He doubts that the computer can handle video recording and project creation simultaneously, although he considers such a test possible.
For JavaScript prototypes, such a configuration may, in the author's assessment, be more convenient than a heavy local editor, especially when a significant portion of the work is performed through cloud models. Browser performance still needs to be tested: a complex 3D scene, screen recording, and the parallel launch of several tools create an additional load.
Conclusion: Arc Engine is suitable for prototypes but does not replace development
In the author's test, Arc Engine showed that the language models used could assemble playable 2D and 3D JavaScript prototypes with the help of pre-prepared prompts and an automated launch. The games included levels, menus, enemies, physical interactions, sound, and basic progression. The review also revealed specific failures.
The main limitation was not a lack of features, but their incomplete consistency. Faulty pipes, enemies coming back to life, inverted controls, repetitive level structures, a boss getting stuck, and overly loud sounds require manual testing. Can Arc Engine immediately release a complete commercial game? Based on the results of this particular test, the author answers rather negatively. Does it help quickly produce a foundation for a prototype? In the author's assessment, yes.
Compare models before you start
The service sets its plans, limits and model catalog. If they differ from this article, contact us so we can update it and record a new review date.
Browse modelsAffiliate link: your price stays the same and the project earns a commission.
