I let my OpenAI Pro subscription lapse two months ago. Claude Max 20x was covering everything. Last week I renewed ChatGPT Pro. Two hundred dollars a month on top of Claude Max.
Anthropic released Opus 4.7 on April 17. Within five days, I went from one AI subscription to two.
Here’s why.
The regression was real — and someone measured it.
AMD engineer Stella Laurenzo analyzed 6,852 Claude Code sessions with 234,760 tool calls. The finding: Opus 4.7 used 80x more API requests and 170x more input tokens to produce measurably worse output than earlier Claude models. Cost impact: 122x more dollars per day on the same workload.
That’s not a vibe. That’s a dataset.
Marginlab, an independent benchmark tracker, paused its degradation detection to reset baselines for the new model. When a monitoring tool stops tracking because the behavior changed too dramatically, that’s a signal.
What the regression looks like day-to-day.
Four specific patterns:
- Reading six files instead of sixty when a task requires broader context
- Asking clarifying questions on choices already specified
- Rewriting entire files when a surgical one-line fix was requested
- Returning summary-level analysis instead of deep investigation
Individually, any one is tolerable. All four together, not small.
There’s a max-reasoning caveat.
Extended thinking mode changes the picture. Opus 4.7 with max reasoning recovers some of that depth. But most production agent workloads don’t run on extended thinking — you’d burn through your weekly quota in an afternoon. At standard settings, the laziness pattern is consistent.
The solution: dual-provider routing.
I now run Claude Code and Codex in parallel. Claude Code wins on visual tooling, prompt caching, and precise surgical edits. Codex wins on web search freshness and architectural analysis.
The dual setup costs $300/month total. What it removes: usage ceiling pressure on both sides. My weekly quota, which I used to hit in two days, now splits cleanly across two providers.
The post covers the specific routing logic, the agent switcher setup that makes context-switching automatic, and why max-reasoning mode doesn’t solve the daily-driver problem.
Read the full post: https://thoughts.jock.pl/p/opus-4-7-codex-comeback-2026
Originally published on Digital Thoughts (Substack).
