# Claude Max 20x — weekly allowance, and what the Opus 5.5 grid cost

Snapshot: 2026-09-25, ~09:30 UTC. Supersedes the 2026-09-06 estimate, which had to *infer* the ceiling. This one doesn't: the app now reports the quota percentage directly, so this uses the same divide-measured-spend-by-reported-percent method as the Codex audit.

## Headline

| | |
| --- | ---: |
| **Weekly allowance (all models)** | **~$2,155** API-equivalent |
| Monthly (× 4) | ~$8,620 — about **43×** the $200 fee |
| Remaining this week (42%) | ~$905 |
| **Opus 5.5 grid, 18 tasks, all attempts** | **$1,267.39 = 58.8% of one week** |

The whole Claude Opus 5.5 benchmark grid cost **just under six-tenths of a single week** of the plan.

## The weekly allowance

The app reports **58% of the weekly all-models quota used**, resetting 2026-10-01 00:59:59 UTC — so the window opened 2026-09-24 00:59:59 UTC. Measuring what was actually consumed inside it and dividing by 0.58 gives the full allowance.

| Component | USD |
| --- | ---: |
| Local Claude Code transcripts, all models, in-window | 648.51 |
| + Oasis Outpost a1 (repo receipt, absent from local) | 91.43 |
| + Jungle Temple a1 | 30.01 |
| + Sekiro high a2 | 11.17 |
| + Courtyard Blossom a1 | 433.73 |
| + Jungle Temple a2 | 34.98 |
| **Corrected window consumption** | **1,249.83** |
| ÷ 0.58 reported | **$2,155 / week** |

Rounding band on the displayed 58% (57.5–58.5%): **$2,136 – $2,174**. That is display precision only, not the full uncertainty — see caveats.

**The five added runs matter.** Local transcripts capture only $756 of the $1,281 of Opus 5.5 work the harness has receipts for — 59%. Those five runs each have a harness receipt, each was git-added inside the window, and none appears in any local session. Using local data alone would have put the allowance at $1,118, nearly half the real figure.

**Internal consistency check.** Opus 5.5 benchmark work is $1,203.37 of the $1,249.83 consumed in this window — **96.3%**. Multiply by the reported 58 points and the benchmark alone accounts for **55.8 of them**. The window is essentially nothing but this grid, which is exactly what the percentage should look like if the measurement is right.

## Per-task cost and share of a week

Costs are the harness's own `api_equivalent_cost` / `api_list_price_equivalent_usd` figures — I did not recompute them.

| Task | Grid effort | Runs | USD | % of week |
| --- | --- | ---: | ---: | ---: |
| Courtyard Blossom Stage 2 | High | 2 | 742.73 | 34.5% |
| Oasis Outpost | High | 2 | 134.98 | 6.3% |
| Sekiro Main Menu | Max | 2 | 105.04 | 4.9% |
| Jungle Temple | High | 2 | 64.99 | 3.0% |
| Carve WebGPU Snowboarding | Max | 1 | 57.68 | 2.7% |
| Space Flight Stage 2 | High | 1 | 33.62 | 1.6% |
| Neon Gunner | Max | 1 | 33.38 | 1.5% |
| Infinite Cathedral | Max | 1 | 28.85 | 1.3% |
| Claw Unlock | Max | 1 | 17.39 | 0.8% |
| Turbofan (CAD-001) | Max | 1 | 14.07 | 0.7% |
| Neo-Gothic Storm City | High | 3 | 8.95 | 0.4% |
| Bluetooth Speaker PCB | High | 1 | 6.47 | 0.3% |
| JRPG Boss Battle | High | 1 | 6.17 | 0.3% |
| Dex Cube Three Moves | Max | 1 | 4.75 | 0.2% |
| Dex Cube Turn | Max | 1 | 2.20 | 0.1% |
| Apple Stem | Max | 1 | 2.11 | 0.1% |
| Eiffel Drawing | Max | 1 | 2.01 | 0.1% |
| Apple Carry | Max | 1 | 2.00 | 0.1% |
| **Total (18 grid tasks)** | | **24** | **1,267.39** | **58.8%** |
| Space Flight Stage 1 (off-grid) | High | 1 | 13.88 | 0.6% |
| **All Opus 5.5 work** | | **25** | **1,281.27** | **59.5%** |

Counting only the runs at the grid's stated effort (dropping the Sekiro high retry) gives $1,256.22, or 58.3% — the choice barely moves the number.

**Courtyard Blossom Stage 2 is 59% of the entire grid's cost.** Its two attempts moved 1.49 billion tokens between them; the failed first attempt alone ($433.73) cost more than the other seventeen tasks combined excluding Oasis. The eight robotics tasks together cost $30.46 — 2.4% of the grid.

Split by window: **$1,203.37 landed in the current window**, $77.90 in the prior one (the 09-22/23 robotics and CAD runs).

## Opus 5.5 rate card

The transcripts carry no cost field, so rates were needed. Solved by least squares from seven independent harness receipts, then confirmed against the card the repo records with `verified_live_pricing: true`:

| $/MTok | in | write 5m | write 1h | cache read | out |
| --- | ---: | ---: | ---: | ---: | ---: |
| **Opus 5.5** | 4.00 | 5.00 | 8.00 | 0.20 | 20.00 |

Write multipliers land exactly on 1.25× and 2.00×. The card reproduces all seven receipts to **0.0000%** error. Opus 5.5 is ~20% cheaper per token than Opus 5 ($5/$25, cache read $0.50).

## Caveats, largest first

- **Mix dependence is the dominant term, and it is large.** The quota meters weighted units, not dollars, so this $2,155 is denominated in an *Opus-5.5-heavy* week (96% of the window). Aligned to the same reset boundary, the Opus-5-dominated week of Aug 20–27 consumed **$4,101** — nearly double this allowance, in a week that may not even have hit 100%. Opus 5.5 therefore burns roughly **2× more quota per API dollar** than Opus 5. Quote $2,155 for Opus-5.5 work; expect closer to $4,000+ for Opus-5 work.
- **The plan string reads "Pro", not Max 20x.** The app reports `plan: "Pro"`. If that is literal rather than a coarse family label, this is a Pro allowance and the Max 20x figure would be higher. The same ambiguity appeared in the Codex audit. Worth confirming in billing.
- **Extra usage: $72.41 already spent.** Overage is currently disabled but has been charged before, so some past weeks ran past their cap — meaning historical weekly totals are not clean ceiling observations.
- **Capture may still be incomplete.** Five missing runs were recoverable because the harness had receipts. Non-benchmark Claude work outside local transcripts (claude.ai web, mobile, cloud) would be invisible and has no receipt. That pushes the allowance **up**, not down.
- **One window, one percentage.** This is a single observation at one point in one window. It is far better evidence than the September 6 inference, but it is not a distribution.

## Method

One record per `message.id` (transcripts hold ~2.09 rows per API response — streaming partials and content-block repeats carry identical input/cache tokens with growing output; 0 input/cache mismatches across 188,104 rows). Input and cache counted once, `output_tokens` taken as the max across a message's rows. Verified rate cards from the harness (`rates_usd_per_million_tokens`). Scripts: `audit.py`, `allow.py`, `hist.py`, `sess.py`.