im running some experiments on this but based on what i have seen on my own personal data - I dont think this is true
"given that Opus 4.7 on Low thinking is strictly better than Opus 4.6 on Medium, etc., etc.”
Opus 4.7 in general is more expensive for similar usage. Now we can argue that is provides better performance all else being equal but I haven’t been able to see that
yeah thats is my biggest issue - im okay with paying 20-30% more but what is the ROI? i dont see an equivalent improvement in performance. Anthropic hasnt published any data around what these improvements are - just some vague “better instruction following"
Yeah this is a bit crazy and not surprising at all.
The limits have always been opaque and you never know when they change.
I started building an open-source local proxy that logs every rate-limit header Claude Code sends.
I am using it to track and get a better sense of the 5h and 7d weekly limits.
Some initial data from 11 observed 5h sessions on Max 20x:
- 5h budget: roughly $120–$280 per window
- 7d budget: roughly $1,300–$1,900
- Separate Sonnet-only 7d budget at ~$150
- 95% of tokens are cache reads. They barely move the meter.
It’s open source so more people can run it and we can figure out the real numbers.
The dropping requirements problem is real. What's helped us is breaking the spec into numbered ACs and having the verification run per-criterion. If AC-3 fails you know exactly what got dropped.
I do it per feature, not per step. Write the AC for the whole feature upfront, then the agent builds against it. I haven't added a spec-validation step before coding but that's a good idea. Catching ambiguity in the spec before the agent runs with it would save a lot of rework
Agreed. The spec file is context. Writing acceptance criteria before you prompt provides the context the agent needs to not go off in the wrong direction. Human leverage just moved up and the plan/spec is the most important step.
Parallelism on top of bad context just gets you more wrong answers faster