Having fun with oh my pi, DeepSeek-V4-Flash, GPT-5.6 Luna and Antigravity CLI

Sun, 02 Aug 2026

alt

Background

I know these are a lot of buzz words, but nevertheless I wanted to share my setup so you can have some fun too! Recently, I’ve been working with omp a “new” coding agent built on top of the very extensible and minimal coding agent pi which also powers OpenClaw. Where pi is minimal, barebones and is left up to the user to customize, omp is batteries included and gives you the agentic experience you’ve come to know and love from Codex and Claude Code while adding even more features on top it.

With the recent announcement that DeepSeek-V4-Flash is no longer in preview, goes head to head with other state-of-the-art models while being extremely cheap, I thought I would give it a try as well.

Adding “vision” to DeepSeek

omp allows you to use your ChatGPT subscription along with a host of providers including search. So I added my DeepSeek API key, logged in via my OpenAI account to use my ChatGPT subscription and got it running. While Flash is extremely good at coding, it does not have multimodal capabilities. This means that if you give it an image and ask it to describe it, it will not be able to do so. It is strictly a text-in, text-out model. While Luna was already cheap, OpenAI slashed the pricing by 80%. This was a good chance to use Luna for vision, give its output to Flash and continue.

This is pretty easy in omp. Just type in /models, choose the model you want to choose for vision and in the bottom nav bar, choose the [vision] option. That’s it. Now any time you paste an image in the chat, Luna would pass the output to DeepSeek or any non-vision model

Screenshot of a dark-themed terminal model selector in omp. The left sidebar lists AI providers, with **openai-codex** selected. The main panel shows six available models: **gpt-5.6-luna**, **gpt-5.6-sol**, **gpt-5.5**, **gpt-5.6-terra**, **gpt-5.4**, and **gpt-5.4-mini**, alongside context sizes, generation speeds, and per-million-token pricing. **gpt-5.6-luna** is highlighted as the current default model and is marked as supporting reasoning and vision, with a 272k context window and pricing of $0.20 input and $1.20 output per million tokens.

Search, search and search

This is where last piece of the puzzle fits in. We’re using one of, if not, the cheapest SOTA and even gave it vision. But what about the newest information out there? For a refresher, when you use any of the 1st party harnesses like Codex or Claude Code and try to query for something recent or something they don’t know, they perform a web search. This is not the case when you use your subscription in another harness (at least not that I’m aware of). The alternative then is to use one of the search providers like Exa, Perplexity, etc. I was not interested in that.

Antigravity CLI to the rescue. Since Google is excellent at search, I thought why not use Antigravity CLI or agy to search and use its output to drive whatever LLM I’m using? The reason you cannot simply log in via Google into omp is because it is against their ToS. So I asked DeepSeek-V4-Flash to create an omp extension to use the agy CLI whenever it wants to use web search. So right now, all the search providers are disabled and omp automatically reaches for agy for search.

Here is the GitHub Gist if you want to use it.

Conclusion

So yeah. That’s the setup I’ve been using since the past few days. DeepSeek-V4-Flash along with GPT-5.6 Luna Max for coding and vision and agy CLI only for search. I’m also trying out herdr, but nothing exciting to report as of yet.