Pro Max subscriber: quality regression cost me days of work and real money

Summary

I'm a solo founder building a cannabis therapeutic intelligence platform (TerpStack). I pay for Pro Max ($200/mo) running Opus 4.6 with max effort. Over the past week, the quality degradation documented in #42796 has cost me real time and real money.

This isn't abstract. Here's what happened.

The concrete harm

I'm running a data pipeline that bulk-matches 5.3M product records across 4 platforms. During planning and execution over the past 5 sessions:

1. Wrong infrastructure sizing advice. I asked multiple times whether my Supabase Micro compute tier ($10/mo, 1GB RAM) would handle the workload. I was told yes, repeatedly. It can't - the t4g.micro instance caps at 500 sustained IOPS, and my table has 21 indexes generating 44 IOPS per row update. Basic capacity planning math that never got done.

1. Surface-level analysis requiring 4+ rounds of pushing. When the pipeline crawled at 25 min per 25K-row chunk, the response was "Supabase IO" hand-waving for multiple sessions. I had to ask 4 separate times in a single session before getting actual research - looking up the AWS instance type IOPS limits, checking index scan counts, calculating write amplification. The research took 5 minutes once actually done. It should have been done on the first ask.

1. Skipped work without explanation. A multi-phase pipeline stopped mid-execution (Phase 7 timed out at 70%) and instead of diagnosing and resuming, the next phases were skipped entirely and the tool moved on to a different platform's data. No explanation, no flag.

1. Days of pipeline time burned. The combination of wrong sizing advice + undiagnosed IO bottleneck + skipped phases means I've lost approximately 4-5 days of pipeline execution time. For a solo founder, that's not recoverable.

The pattern

Every response defaults to shallow. I have to push repeatedly to get analysis that should be the starting point:

• "Supabase IO is the bottleneck" (no research into what that means)
• "Consider upgrading compute" (no pricing research, no alternatives)
• "I don't have a good explanation for why my quality varies" (when asked directly about the degradation)
• Generic apologies instead of action

This matches exactly what #42796 documents - the Read:Edit ratio collapse, the "simplest fix" mentality, the shallow thinking that requires human pressure to deepen. I'm paying for max effort on Opus 4.6 and getting what feels like medium effort on Sonnet.

What I've tried

• /effort max - set from the start of every session
• Model set to Opus 4.6 with 1M context
• Detailed CLAUDE.md with project context

None of these restore the quality I had in January-February 2026.

What I'm asking

1. Don't close this as "completed" with settings suggestions. The settings don't fix it.
2. The system prompt change in v2.1.64 ("try the simplest approach first") is actively harmful for complex engineering work. It should be removed or configurable.
3. Adaptive thinking under-allocates on complex multi-step tasks. This was acknowledged on HN as being investigated - what's the status?
4. If the model is being routed to a cheaper tier despite my subscription, that needs to be fixed and disclosed.

Environment

• Claude Code v2.1.100
• Pro Max plan
• Opus 4.6 (1M context)
• Effort: max
• macOS, terminal CLI