Trying to Code with GPT-5.6

On July 11, 2026, OpenAI announced three types of GPT-5.6.

• Sol: The smartest model
• Terra: A smart model for work
• Luna: A fast model for everyday use

So, I immediately tried using the smartest GPT-5.6 Sol with Codex.

Slower than expected but faster than 5.5

Well, I’m not sure. When using 5.5, I had to wait a very long time, but I felt it was faster than that.
However, it’s nothing compared to DeepSeekV4 Flash.

The feature I implemented this time is the article translation function added to this blog app.
A function that translates articles using AI asynchronously after users write and publish them.
AI uses Cloudflare Workers AI’s Quen3.
It seems that asynchronous job execution can be done with Rails’ standard functionality. It’s implemented nicely.

After running for an hour, I almost used up the weekly limit

n
Sol created most of the features after running for an hour, but at that point I had almost reached the weekly limit, so I downgraded to Luna.
In the Plus plan, Sol can’t be used except for very difficult problems.

The user experience is OpenCode + DeepSeekV4 Flash > Codex + GPT-5.6

Personally, I felt OpenCode + DeepSeekV4 Flash was better.
First of all, it’s much faster. It responds immediately, and the work speed is quite different.
Moreover, its implementation and design skills are quite high, so it can quickly and correctly create most things.

There seems to be a difference between GPT-5.6’s intelligence and the OpenCode + DeepSeekV4 Flash workflow of “quickly trying various options to find solutions”.
OpenCode + DeepSeekV4 Flash almost felt like a human, checking production server logs with kamal and even inspecting the contents of the production server’s database for troubleshooting.

It might be interesting to collaborate with it

What I’m currently thinking about is whether it would be good to give OpenCode the skill of “consulting Codex”.
Since Codex can be called as an external command, it would be great if we could call Codex from OpenCode’s skills for difficult designs or bugs that can’t be fixed.

$ codex --model "gpt-5.6-luna" exec "こんにちは"

OpenAI Codex v0.144.1
--------
workdir: /home/hiroe/src/tech_notes
model: gpt-5.6-luna
provider: openai
approval: never
sandbox: workspace-write [workdir, /tmp, $TMPDIR]
reasoning effort: medium
reasoning summaries: none
session id: xxxxxxxxx-39b3-xxxx-aebc-xxxxxxxxxxxxxxxx
--------
user
こんにちは
codex
こんにちは！今日はどのようなお手伝いをしましょうか？
tokens used
5,647

Since tomorrow is the weekend, I’ll try it if I have time.