Hello,
I’m opening this new post to discuss the different agents for developing external integrations and I’ll take the opportunity to respond to @pierre-gilles’s post 
:
How would DeepSeek V4 Flash compare to Claude Opus in enhanced mode, considering that the limit of DeepSeek V4 Flash with opencode is visibly higher than with Claude Opus in enhanced mode?
Claude code is also easier to use as it’s directly in web mode, even if opencode allows it, but you need some knowledge to make it accessible from anywhere and secure it.
The goal would be to develop and/or maintain external integrations at a lower cost 
I understand that $20 per month isn’t much for everything Claude allows regarding external integrations, but it’s still a small budget to create/maintain external integrations for which we don’t make money. 
And the limits are melting away, even more so after August 19th when the Claude promotion ends
There are three things when you develop:
- the LLM model and clearly, frontier models are above the others. DeepSeek v4 or Kimi K3 are getting close
- the system prompt and the harness around it (Claude Code or OpenCode)
- The context (AGENTS.md and your skills)
I recommend using Claude Code with Opus/Fable or Codex with GPT 5.6 Sol for analysis and planning (which you can save in a markdown file). But you can use cheaper models for implementation.
There are subscriptions like Devin that allow you to test multiple models.
Or OpenRouter/OrcaRouter which are proxies to all models (but you pay per use).
I wouldn’t be able to recommend anything other than Claude today
Even for Codex, the feedback I’ve seen on this forum hasn’t convinced me. For example, regarding the Apple TV integration, the maintainer spent several iterations trying to get the code to work with Codex. In the end, he started from scratch with Claude Code, which generated a functional implementation on the first try.
In my opinion, Claude stands out in two areas:
- The models are, today, among the best for development.
- The « Claude Code » tool, thanks to the prompts and the environment designed by Anthropic, offers a development experience that is close to that of a real software engineer at our side. It seems to naturally understand the right steps to follow, runs the tests, restarts them when necessary, validates the behavior in real conditions and iterates until a reliable result is obtained, rather than just generating code.
After all, this field is evolving at a breakneck pace. In two weeks, my opinion might have completely changed.
But as of now, not using Claude mostly gives me the impression of wasting a lot of time fixing code generated by less performant or less well-orchestrated models. The cost of the subscription is largely offset by the time saved.
Thanks for the feedback
I saw that you recommended using Opus 5 in high for the video you made for the SMTP integration
Do you think you recommend this to start the development of the integration but for debugging and adding features, switching to high sonnet or low or medium Opus 5 to save tokens is something you would consider?
Yes, Sonnet 5 might be enough, it’s up to you to do your tests and form your own opinion
I rarely, rarely use Sonnet 5 on my end, but sometimes it’s enough.
Hi @prohand. I already use Cline in pro mode because it allows me to use local LLMs with Ollama and I’m currently testing Cline Pass which is the same as OpenCode Go. It’s very recent, just like OpenCode Go, coincidence?
I use it for many different projects and it works very well. I use the Qwen 3.7 Max model for reasoning and project architecture. It’s the brain I use exclusively in Plan mode. For the rest, it’s DeepSeek v4 pro in Act mode and also in Plan mode if the architecture is already well defined. You can also use a free version of DeepSeek v4 flash for small modifications and tests.
With Cline, you can’t do it in the browser for now. However, it’s compatible with IDE and terminal.
