Some thoughts on Codex vs Claude

With Opus 5.5 I have switched to using Claude as my main driver and here are current thoughts to improve the state of Codex. Disclaimer: I have both a Anthropic and OpenAI $100 subscription so im unbiased. Raise the limit for the double priced context limits to 400k tokens, it doesn't have to be 1M but 258K is woefully little for large codebase especially when working in singleton + critic workflows Roll out the new experimental compaction tool as a stable version - Right now claude has a memory tool that helps sessions remember things and intent on long conversion through compactions. I think the new experimental compaction is the step in the same direction. Release an Astra light (Astra is great) or lower the Astra price. Sol being cheap is moot because you'll burn 10x more tokens than Astra due to it trying to over validate and over-engineer a feature. This has severe downstream token costs because now future agents need to read 2x-4x the tokens for the same feature. Fix the subagent polling issue, it should be event based not time based (or at least allow me to change the pulse) - there's no reason why I need updates every 1 minute hemorrhaging tokens. Claude has a much better harness with subagents being able to wake the main thread when it's done. The worst part is this is baked into the system so I can't even change it with my agents.md Have a better auto approve system: this is kinda niche but I notice the auto approver is very pedantic sometimes and it doesn't ping me to approve it manually. I watch agents do some weird detours instead of just asking for approval.