How much does it really cost to use an LLM to write code? Not in abstract terms - tokens per million, list prices - but on a concrete, measurable task with a verifiable result.
We tried to answer that. Using AIDeskPro as a gateway, we tracked costs and execution times for an identical prompt across 5 models and 3 different harnesses. The task: create the Pong game in a single HTML file.
The prompt
Create a complete, playable Pong game in a single HTML file called exactly pong.html, saved in the current directory. Requirements: HTML5 canvas; vanilla JavaScript with no external dependencies; two paddles; a ball that bounces off walls and paddles; on-screen score for both players; left paddle controlled by keyboard (arrow keys up/down or W/S) and right paddle with a simple AI. Actually write the file to disk as pong.html. When done, reply with a single word: DONE.
A real task, not a synthetic benchmark. The model must reason, write working code and save the file to disk. No tricks.
The harnesses
We tested three CLI tools for coding agents:
| Harness | Version |
|---|---|
| Codex | 0.144.4 |
| ClaudeCode | 2.1.217 |
| Opencode | 1.18.4 |
The results
| Model | Codex | ClaudeCode | Opencode |
|---|---|---|---|
| GPT-5.6 Terra | €0.0577 / 12.8 s | €0.0759 / 9.9 s | €0.0588 / 20.3 s |
| Claude Sonnet 5 | €0.2050 / 69.4 s | €0.2569 / 17.2 s | €0.0796 / 28.5 s |
| Gemini 3.6 Flash | €0.0887 / 26.5 s | €0.1122 / 18.1 s | €0.1020 / 28.0 s |
| GLM-5.2 | €0.0459 / 38.3 s | €0.1037 / 67.4 s | €0.0461 / 37.2 s |
| Kimi K3 | €0.0749 / 96.2 s | €0.0664 / 54.2 s | €0.0538 / 91.7 s |
| Total per harness | €0.472 / 243 s | €0.615 / 167 s | €0.340 / 206 s |
Every cell is clickable: you can play the Pong produced by that model/harness combination directly.
What stands out
GPT-5.6 Terra with codex is the fastest combination: 12.8 seconds, €0.0577. Hard to beat for a standard coding task.
GLM-5.2 with codex or opencode has the best price/quality ratio: €0.0459-0.0461, reasonable time, quality output. GLM-5.2 is served by AIDeskPro via Scaleway, with European endpoint and zero data retention - making it also the safest choice for those working with business data.
Claude Sonnet 5 shows high variance across harnesses: with codex it costs €0.20, with opencode it drops below €0.08. The harness matters.
Kimi K3 is the slowest overall: over 90 seconds with codex and opencode. It is not a model designed for low latency on simple tasks.
opencode is the most cost-effective harness overall: €0.340 vs €0.615 for Claude Code. At scale, the difference is significant.
Why AIDeskPro as a gateway?
Tracking real costs per token and per call is not straightforward when using multiple models and multiple tools. AIDeskPro centralises access to all major LLMs and records costs and consumption in real time, for every session and every user.
This benchmark came about exactly this way: an internal team used AIDeskPro to run a comparative test and ended up with a clean, reproducible measurement with real numbers.
If you want to do the same in your company - compare models on real tasks, control costs, understand which combination works best for your workflow - that is exactly what AIDeskPro is built for.