Why I'm Testing OpenCode + GPT-5.3 Codex
Feb 24, 2026
For the last year my LLM of choice has been Claude Code heavily. The built-in subagent functionality is a huge part of why I prefer it. Being able to spin up a few focused personas to help with planning, critique, and implementation is the thing that made it feel like a real workflow instead of a chat session.
So why test OpenCode with GPT-5.3 Codex now?
1) I want to reproduce the subagent workflow
This is the main question. OpenCode supports agents and subagents too, and I’m testing whether it can reproduce the flow I had with Claude Code. If it can, then I get the best part of Claude Code without being locked into a single client. OpenCode’s agents doc is here if you’re curious: opencode.ai/docs/agents
2) The OpenCode UI is a LOT better
OpenCode’s local web UI is simpler and more usable for my workflow, and it works cleanly in WSL2. That last part is not optional for me. The docs are explicit about WSL2 support and setup here: opencode.ai/docs/windows-wsl. And the web UI itself: opencode.ai/docs/web
3) I don’t like the way Anthropic handled OpenCode
OpenCode users were using Claude Pro/Max credentials for Claude Code, and Anthropic blocked it. I understand the business reasons, but I don’t like it, and I don’t want to build my workflow on something that can get turned off when a vendor decides they don’t like the client. There are public GitHub issues documenting the block:
- https://github.com/anomalyco/opencode/issues/10982
- https://github.com/anomalyco/opencode/issues/10956
OpenCode’s provider docs even call out that Claude Pro/Max credentials are not officially supported: opencode.ai/docs/providers
4) GPT-5.3 Codex looks competitive right now
There’s a Terminal-Bench 2.0 leaderboard that has GPT-5.3 Codex at the top, with Claude Opus 4.6 close behind. That’s not the only benchmark I care about, but it’s a meaningful signal for coding performance: Terminal-Bench 2.0 leaderboard
Also, Simon Willison notes that GPT-5.3 Codex dropped the same day as Opus 4.6, and is currently only available in the Codex app (not API yet). That’s a big constraint and part of why this is still an experiment for me: two new models
5) Cheaper
Anthropic’s pricing for Opus 4.6 and Sonnet 4.6 is here:
OpenAI’s pricing page doesn’t list GPT-5.3 yet, but GPT-5.2 Codex was noticeably cheaper and I don’t see 5.3 massively jumping in price. Anthropic’s pricing is a bit insance, to be honest.
Now, of course, when you’re on one of the $200/mo plans for either that’s kind of a moot point. It’s essentially unlimited. But I don’t think that free ride can last forever. OpenAI and Anthropic have got to be losing money on these $200/mo subs, so at some point they’re going to have to end them, or restrict them, or otherwise kneecap them.
So what am I actually doing?
I’m running OpenCode as my daily driver for a bit and seeing if it can recreate my subagent workflow. If it can, I keep the workflow I like, with a better UI and fewer platform risks. If it can’t, I go back to Claude Code and keep paying the tax for the things it does better.
This is all a moving target right now. But if you’re curious, the OpenCode docs are a good place to start: opencode.ai/docs