At our level of use, and that is building indicators, building strategies, building new tools and perspectives of the market with artificial intelligence, it’s been my experience that Claude 4.6 is superior to 4.7 and 4.8.
I’ve built incredible tools that I’ve shown in this forum, through my posting here that would not have been possible with Claude 4.7 or 4.8 primarily due to the fact that 4.7 and 4.8, actively discourage you from incredible sophistication, it’s through their verbiage and their communication that this is hard, this is not possible and such. Additionally, claude 4.7 and 4.8, make more semantic mistakes in their code, they assume more, and this assumption is what has been trained into them as an improvement, but it is debilitating when it comes to coding.
Claude 4.6 simply built anything that was required of it straight. So if you find yourself coming up against a brick wall, use Claude 4.6.
I only use free, for now, versions of Claude 4.6, Qwen 3.7, ChatGPT 5.5, Gemini 3.5Flash. Gemini, ChatGPT are good for research, planning and debugging, Claude coding, Qwen for making changes to already structured code/debugging. When 4.7 was available, I had similar experience as you, 4.6 was much better following instructions.
Greatly limited number of tokens works for me as otherwise I would be vibe coding all the time, however, I want to try using a coding harness like Pi, VS Code/Aider or Claude Code which means I’ll have to pay for API at some point.
You don’t need to pay for API usage with Pi. You can use a ChatGPT subscription with it. I use Pi and have developed a custom system around it and wouldn’t suggest it unless you want to use something that is highly customizable. I think most people are good with sticking with Claude Code or Codex.