NEW DeepSeek v4 Flash Built These Games... for $1!
We just ran DeepSeek's new V4 Flash through our standard game dev gauntlet, and the results force a hard look at how you spend your API budget. This is the 731 model, a lightweight release with 284 billion total parameters but only 13 billion active per token. That ratio matters because it keeps reasoning consistent without the compute cost of the trillion-parameter flagships.
We had it build 11 mini projects, everything from a Helms Deep simulator to a mechanical watch render. Total cost: under $1. That covered 200 million tokens and 868 API requests. For context, the same workload ran us about $30 on GPT-5.6 Sol and $57 on Fable 5. If you are prototyping or running auxiliary agents, that price difference changes your workflow math completely.
Where it actually delivers
The model handled tasks we normally reserve for the big boys. It built a working Jabberwocky game with progression logic, a petri dish predator-prey simulation with accurate population graphs, and a low-poly Minecraft clone with crafting. The pathfinding in the office life sim was janky (workers walked through fridges), but the blackout behavior was a new feature we had not seen from other models.
Its planning ability surprised us. We prompted it to orchestrate sub agents for the builds, and it defined scope, supervised execution, and caught bugs in the sub agents' code midway through. That is not typical lightweight behavior. It also handled the Chinese architecture assembly test well enough to complete a roof, something even frontier models often fail to finish.
Where you still need the heavyweights
Do not let the benchmark charts fool you. DeepSeek's marketing numbers look better than GLM 5.2, but remember that GLM runs far more total and active parameters. When we compared outputs directly, the gaps showed up in polish and logic.
Claude Opus 5 still owns the intricate tasks. Its Helms Deep had working drawbridge logic and better rain effects. Its mechanical watch simulator showed actual gear geometry with layered depth, while V4 Flash produced a simpler rotating assembly. Kimi K3 remains the defending champion for Jabberwocky, with superior game progression and texture details in the age erosion tests.
V4 Flash also whiffed on the universe simulator, producing dark, pixelated planets that looked nothing like the standard outputs we see from models with this in their training data. The trebuchet physics worked, but the visuals were rough compared to Opus 5's medieval warwolf build.
How to actually use it
Treat V4 Flash like a very smart intern, not a senior architect. Use it for research tasks, simple coding iterations, and mundane automation. If you run a multi-agent setup, let the trillion-parameter models handle the hard coordination while V4 Flash cranks out the smaller components and pull requests.
We ran it through Codex on a VPS using the default high reasoning effort. Do not max out the reasoning. Lightweight models overthink when you push them, and V4 Flash performs best when you give it clear instructions and let it execute. It follows directions well, which makes it perfect for auxiliary roles.
The bottom line: if you are bootstrapping a project and cannot justify $20 to $50 per build session, V4 Flash gets you 80% of the way there for literal pocket change. Just do not expect the wow factor or the intricate logic of the flagships. Match the tool to the task, and your wallet will thank you.