What are your token usage experience with Opus 4.7?

[pySilver]

Hi!

I'm curious what are your experiences with Opus 4.7 with Claude Code? I use x20 Max subscription in a moderate way. Single day of work: 2 medium-sized plans at 21% of my weekly budget. I'm kinda not impressed. Last plan ate about 6% of usage on moderately quick planning and execution:

plan:     /workspace/docs/plans/2026-04-17-asyncio-to-anyio-migration.md
branch:   asyncio-to-anyio-migration
mode:     full
duration: 2h43m
changes:  63 files (+634/-1070 lines)

I used xhigh during planning and high effort on execution:

--task-model=:high --review-model=:high

I do understand that it's apples vs oranges, so my goal here is only to collect your experience:

• Are you happy with Opus 4.7?
• Do you plan to switch to 4.6 or 4.5 instead?
• Maybe older model for execution and newest for planning?

TBH - I tried switching to 4.5 a few days ago and it was OK.

P.S. Worth reading: https://claude.com/blog/best-practices-for-using-claude-opus-4-7-with-claude-code

[pySilver]

Post on reddit on this: https://www.reddit.com/r/ClaudeCode/comments/1sogcki/best_token_tradeoff_for_detailedplan_agentic/

[pySilver]

I've run an experiment with a highly detailed plan: 10 tasks with 83 subtasks in total. Plan was created with Opus 4.7:xhigh; On every run I preserve code outside of the git repo, so next run can't cheat by discovering solutions made in previous runs. Sadly, I'm unable to track token spend other than tracking my weekly limit usage as coding sessions are executed in dockerized environment.

1st run

Coding model: Opus 4.7:medium

Review model: Opus 4.7:xhigh

Initial multi-agent internal reviews: 4 iterations

Multimodal Codex vs Opus review took: 10 iterations

Final internal multi-agent reviews: 1 iterations

Results: Weekly limit on x20 account spend: 8% (implementation + reviews)

2nd run

Coding model: Opus 4.5:high

Review model: Opus 4.7:xhigh

Initial multi-agent internal reviews: 2 iterations

Multimodal Codex vs Opus review took: 3 iterations

Final internal multi-agent reviews: 1 iterations

Results: Weekly limit on x20 account spend: 1% (implementation) + 4% (reviews)

3rd run

Coding model: Opus 4.5:high

Review model: Opus 4.5:high

Initial multi-agent internal reviews: 2 iterations

Multimodal Codex vs Opus review took: 6 iterations

Final internal multi-agent reviews: 1 iterations

Results: Weekly limit on x20 account spend: 1% (implementation) + 4% (reviews)

4th run (pending)

Coding model: Opus 4.6:high

Review model: Opus 4.6:high

Initial multi-agent internal reviews: X iterations

Multimodal Codex vs Opus review took: X iterations

Final internal multi-agent reviews: X iterations

Results: Weekly limit on x20 account spend: X% (implementation) + X% (reviews)

[pySilver]

Plan execution speed (includes tests, but that ~ constant time)

47m-47xh: 4h
45h-47xh: 2.5h
45h-45h: 3.5h
46h-46h: pending

[pySilver]

Full set of setups:

I - Implementation phase
R - Review phase

47m-47m

Initial reviews: 2 iterations
Multimodal reviews: x iterations
Final reviews: 4 iterations
Results: Weekly limit on x20 account spend: 3% (I2% + 1R)
Time taken: 2h

47m-47xh

Initial reviews: 4 iterations
Multimodal reviews: 10 iterations
Final reviews: 1 iterations
Results: Weekly limit on x20 account spend: 8% (I + R)
Time taken: 4h

46h-46h

Initial reviews: 2 iterations
Multimodal reviews: 7 iterations
Final reviews: 1 iterations
Results: Weekly limit on x20 account spend: 7% (I2% + R5%)
Time taken: 3h40m

45h-47xh

Initial reviews: 2 iterations
Multimodal reviews: 3 iterations
Final reviews: 1 iterations
Results: Weekly limit on x20 account spend: 5% (I1% + R4%)
Time taken: 2h30m

45h-45h

Initial reviews: 2 iterations
Multimodal reviews: 6 iterations
Final reviews: 1 iterations
Results: Weekly limit on x20 account spend: 5% (I1% + R4%)
Time taken: 3h30m

[pySilver]

Benchmark results

Five model runs produced five branches against the same implementation plan. Two review passes were performed: Codex ranked them on how completely they satisfied the plan, then Opus-4.7:xhigh re-graded the same branches on correctness (race-safety, idempotency, signal hygiene) and validated Codex's specific claims.

Ranking shift across passes

Branch | Pass 1 (completeness) | Pass 2 (correctness) | Δ -- | -- | -- | -- 47m-47xh | 9.0 | 6.5 | −2.5 47m-47m | 9.5 | 6.0 | −3.5 45h-45h | 8.5 | 6.0 | −2.5 46h-46h | 9.0 | 6.0 | −3.0 45h-47xh | 9.0 | 5.5 | −3.5

What each branch got right and wrong

• 47m-47xh — best overall. Strongest on the hardest parts: concurrency guards, locked re-reads, and correct retry scheduling. Weaker on error-mapping consistency in one subsystem. Wins both rankings.
• 47m-47m — most complete, not most correct. Satisfied the plan most thoroughly and had the cleanest error handling in one area, but introduced a retry-scheduling bug that can disrupt in-flight work. Biggest drop between passes.
• 45h-45h — least complete, held up best. Missed several checklist items but was the only branch to handle signal registration correctly and got retry semantics right. Smallest drop. Its weakness is a narrow scheduling filter that leaves some failures stuck.
• 46h-46h — mechanically complete, quietly buggy. Avoided one race the others had, but introduced a worse one by overwriting active scheduling state unconditionally.
• 45h-47xh — broad but shallow. Good coverage across subsystems, weakest concurrency guards on the critical path. Lowest correctness score.

What the two passes revealed about the models

• Plan-coverage ≠ code quality. The most complete branch dropped 3.5 points on correctness; the least complete dropped 2.5. Finishing the checklist did not predict getting the hard parts right.
• The plan was under-specified on correctness concerns. Concurrency primitives, idempotency, and signal hygiene weren't mandated, so branches that stuck closest to the letter of the plan scored lower when those dimensions were graded.
• Verification caught real errors in the first review. When Opus re-examined Codex's specific per-branch claims, roughly half needed revision — some bugs were real but described incorrectly, some weren't there, and in a couple of cases a different bug was in the same place. The headline ranking mostly survived; the justifications often didn't.
• No branch was merge-ready. Every model left at least one correctness gap the plan hadn't explicitly required it to close.

Takeaway for the benchmark

Running the same plan across models produces meaningfully different artifacts, and the ranking you get depends heavily on what you grade for. A single review pass — even from a strong model — overstated quality across the board. The two-pass setup (independent review, then cross-model validation) was what made the residual-risk picture legible.