Low‑Cost Game Development with OpenCode AI: How Inference Routing Drives a Godot Shootout

Connecting OpenCode to DigitalOcean

Created an inference router named game-designer with four routes:

design : primary model GLM-5.2 with backups Deepseek V4 Pro, Kimi K2.6, MiMo V2.5 Pro.

repo-qa : models Gemma 4, MiMo V2.5, Arcee Trinity Large Thinking, Qwen 3.5 397B A17B.

chore : default Gemma 4 with backups Qwen3 Coder Flash, Ministral 3 14B, MiMo V2.5.

debug : models GLM-5.2, Deepseek 3.2, MiniMax M2.5, Kimi K2.6.

Connecting required a single /connect command, OAuth login, then selecting the router via /models → router:game-designer.

Model usage distribution

During the build the router dispatched 596 tasks into five buckets: chore (335 tasks, 56.2%), design (207, 34.7%), debug (41, 6.9%), repo‑qa (11, 1.8%) and 2 fall‑backs (0.3%).

Only three models handled almost all traffic:

MiMo V2.5 – ~56% of calls

GLM‑5.2 – ~42% of calls

Gemma 4 – ~2% of calls

GLM‑5.2’s share matches the combined design and debug routes (34.7% + 6.9% ≈ 41.6%). The chore route defaulted to Gemma 4, but the router chose MiMo V2.5 because its cost‑performance ratio was higher.

Cost, latency and token usage

Assistant turns: 83; user turns: 17 (≈ 5:1 ratio). Net generation time: ~74.5 minutes (2.1 s for a tiny fix up to 529.6 s for a full README). Router latency was stable at 0.27–0.30 s per request (measured on 2023‑06‑25 for game-designer).

Token usage: ~4.1 M tokens, costing $8.25 total:

MiMo V2.5 – $0.65

GLM‑5.2 – $7.56

Gemma 4 – $0.04

Running all 596 tasks on a frontier model would have cost ≈ $123; using only GLM‑5.2 would have cost ≈ $18.04.

Human intervention points

1. Real code bug (agent solved autonomously)

The ball did not launch because the input event never reached _unhandled_input. A full‑screen ColorRect node had MOUSE_FILTER_STOP, swallowing mouse events. Changing the filter to ignore fixed the issue with a single line of code.

Note: The agent diagnosed the root cause without explicit guidance.

2. Spatial ambiguity (human clarified intent)

Prompt: move the goal “about 1/3 of the way down the screen”. The agent interpreted “down from the top” and moved the goal upward. After clarification “1/3 from the bottom”, the next attempt placed the goal correctly.

Lesson: Ambiguous spatial language adds dialogue turns; precise prompts reduce effort.

3. Missing external knowledge (agent needed real API details)

When asked to generate a DigitalOcean serverless image‑generation script, the agent fabricated endpoint URLs it did not know. It flagged uncertainty, requested the correct endpoint and token, and produced a generic OpenAI‑compatible script with placeholders.

Core limitation: Models without retrieval cannot reliably provide up‑to‑date external API details.

When to use routing vs. frontier models

Routing suitable for: routine development tasks—high‑volume, low‑risk edits, design work, and simple code‑base queries. The cost data ($8.25 vs. $123) demonstrates the economic advantage.

Frontier models required for: tasks that depend on the latest external system knowledge or need multi‑step, long‑horizon architectural reasoning. Mid‑tier models may hallucinate API details, leading to broken code.

Configuration and reproducibility

All steps can be reproduced with a DigitalOcean account, OpenCode, and a few minutes of setup. Source code is open‑source at:

https://github.com/Jameshskelton/PK_Shootout

Create a model access key in DigitalOcean Inference → Serverless Inference.

Optionally create the game-designer router with the four routes described above.

Install OpenCode, start it, and run /connect followed by OAuth login.

Select the router with /models → router:game-designer and begin coding. /connect Example Python client (compatible with OpenAI, cURL, Gradient Python SDK, LangChain, LlamaIndex):

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://inference.do-ai.run/v1/",
    api_key=os.getenv("MODEL_ACCESS_KEY"),
)

resp = client.chat.completions.create(
    model="game-designer",  # router name or base model ID
    messages=[{"role": "user", "content": "Summarize the repository architecture."}],
)
print(resp.choices[0].message.content)