DeepSeek Harness Builds an Arcade Game from a Single Prompt in 2 Hours 43 Minutes on a $10 Subscription

Usually, "generated an app from a single prompt" means a pretty demo that falls apart on first contact with reality. The author of this story decided to test what happens if you give an inexpensive model a proper agentic environment and a genuinely big task.

What Exactly Happened

The idea is simple: a browser game for practicing typing speed. Words fly across the screen, the player types them, shoots enemies, earns points and tries not to lose lives. The author first developed the spec in regular ChatGPT: mechanics, screens, constraints, checks. The finished document went into DeepSeek Harness, an open agentic environment from DeepSeek built on the Cordis architecture with the principle "Everything is a Plugin".

The model was DeepSeek V4.1 Flash, accessed through an OpenCode Go subscription at $10 per month. The agent built the first version in 2 hours 43 minutes: 183 steps, Agent Team mode with subagents, command execution, tests, rework. The result: a game loop, enemy waves, combos, Canvas 2D graphics with particles and explosions, sound via the Web Audio API, Russian and English dictionaries, and a run history in localStorage.

An important caveat from the author: "from a single prompt" does not mean "in a single model response". It is one goal, within which the agent planned, wrote and fixed things by itself.

The Numbers Worth Looking At

The interface showed 35,703,377 tokens. It sounds scary, but 35.2 million of them were cached input, uncached input was 334 thousand, and the model's actual output was about 159 thousand. Cache hit: 99%. In long tasks the agent constantly rereads the project context and its own earlier results, and if the provider reuses the cache efficiently, this costs noticeably less.

The task cost roughly 15% of the subscription's five-hour limit. The author honestly warns that "$100 of usage for $10" is a calculated volume at the service's rates, not a balance you can spend on anything. There are five-hour, weekly and monthly windows. There are also Peak and Off-Peak modes with different token prices: for example, $0.15 versus $0.30 per million input tokens. The author even wrote a small plugin that shows the current mode in the Harness interface, so big experiments can be launched in the cheap window.

A separate mention goes to a post by an OpenCode developer about the possibility of getting $300 of DeepSeek usage for $10. For now this is only a public statement about potential, not a term of the plan.

Where a Human Had to Step In

For the second task, the agent added a server: Telegram authentication, stats sync, friends, a leaderboard, PostgreSQL and a support bot. The game became Typer Cool. And that is where real problems began:

• mobile layout that looked fine in Chrome DevTools fell apart on a real iPhone Safari;
• the generated Docker Compose brought up a second Caddy and conflicted with the one already running on ports 80 and 443;
• Telegram avatars did not load because of an overly strict format check (bad_type), even though tests with stubs passed;
• direct invitation links returned 404 on Netlify until the SPA routing rules were fixed.

Practical Takeaway

For developers and founders there are four lessons here. First, the quality of the task definition decides everything: discuss the idea and build the spec first, then hand it to the agent. Second, the environment matters as much as the model: files, subagents, tests and long context change the outcome. Third, caching and pricing should be studied as carefully as benchmarks. And fourth, an agent does not remove the need to verify on real infrastructure, especially when users, data and security are involved.

The limitations are worth remembering too. This is a single experiment by a single author, with no comparison against Codex or Claude Code on identical tasks, and the 99% cache hit applies to one specific session. Harness is under active development, interfaces may change, and so may pricing. Hosting, a domain and a VPS are not included in the subscription cost. But as a benchmark for the price and capabilities of long agentic development on an open model, the case is telling.