Measured capacity of subscription usage windows, for sizing eval arms and routines against the 5-hour window.
Each entry drives one window from 0% to the session limit with a known workload and records raw tokens from peek plus the API-equivalent cost at list price. The Usage page converts tokens to plan percent from these numbers. Cache-field behavior: Prompt caching.
Measured 2026-09-28.
One Team Premium 5-hour window held $262.04 of API-equivalent Opus 5.5 spend at list price, reached in ~48 minutes at this fan-out (≈ $5.5/min). The all-models weekly limit fell from 100% to 84%: one window is 16% of the week, ~6.25 windows a week. Fable was not in the run; it may use at most half the weekly limit.
| Item | Value |
|---|---|
| Window | Opened 19:40 Berlin (first arm launch 19:40:22); the limit message named the reset 00:40, exactly 5 h later. |
| Workload | /skillroutine-eval railroad-review preset (g2 bootstrap, smine 1.1.4..1.2.1), arm claude-opus-5-5[1m]:high, lanes and station opus:medium. All 106 agents ran claude-opus-5-5. |
| Telemetry | peek 1.2.10, central instance (control port 42442). CLI total_cost_usd and OTEL cost_usd agree. |
| Load | Wall time | Cost (CLI/OTEL) |
|---|---|---|
Arm 1, 2026-09-28-a34281 (aborted before round 1) | 19:40:22–19:41:48 | $1.06 |
Arm 2, 2026-09-28-a34281-2 (dispatcher + 105 subagents, killed by the limit) | 20:16:58–21:04:41 | $260.98 |
| Window, 0% → limit | $262.04 |
Main session plus all subagents, from peek.
| Category | Tokens | Opus 5.5 rate | Cost | Share |
|---|---|---|---|---|
| Cache write (all 1h TTL, subagents included) | 17,216,269 | $8 / MTok | $137.73 | 53% |
| Cache read | 390,238,659 | $0.20 / MTok (0.05×) | $78.05 | 30% |
| Output | 2,137,757 | $20 / MTok | $42.76 | 17% |
| Input | 4,834 | $4 / MTok | $0.02 | 0% |
| Recomputed | $258.56 |
The recomputation is 0.9% below the reported $260.98, most likely calls in flight at the kill.
Opus 5.5 list prices (platform.claude.com pricing, fetched 2026-09-28), per MTok: input $4, 5m write $5, 1h write $8, cache read $0.20, output $20; 1M context at standard rates. The 0.05× read rate differs from other models' 0.1× and Fable/Mythos 5.1's 0.025×.
| Role | Agents | Write | Read | Output | Total |
|---|---|---|---|---|---|
| Code-style lanes | 8 | $25.53 | $13.88 | $10.13 | $49.53 |
| Verifiers | 40 | $24.98 | $10.08 | $9.13 | $44.19 |
| Tests lanes | 8 | $21.73 | $16.81 | $5.13 | $43.67 |
| Correctness lanes | 8 | $20.30 | $14.69 | $5.72 | $40.71 |
| Refuters | 25 | $20.77 | $11.65 | $6.58 | $39.00 |
| Security lanes | 8 | $18.76 | $9.17 | $3.98 | $31.91 |
| Root dispatcher | 1 | $1.39 | $1.57 | $0.73 | $3.69 |
| Checklist agent | 1 | $2.43 | $0.19 | $0.88 | $3.49 |
| Groupers | 5 | $1.84 | $0.01 | $0.49 | $2.34 |
| Station | 2 | — | — | — | $0 (never ran) |
Lanes are 64% of the spend. Writes, not reads, are the largest line: every agent writes its prompt at the 1h TTL (2× input), and Opus 5.5 reads are cheap. Valid only for Opus 5.5 at this fan-out shape. Earlier Team Premium estimate: $260–290 (2026-09-05).
The Usage page (internal/budget/plan.go) derives plan percent from this window alone; no dollar value takes part. The unit is one Opus 5.5 output token; every other token is weighted by its list-price ratio to it, cache writes at the 1-hour rate.
| Model | Output | Input | Cache write | Cache read |
|---|---|---|---|---|
| Opus 5.5 | 1 | 0.2 | 0.4 | 0.01 |
| Fable 5.1 | 2.5 | 0.5 | 1 | 0.0125 |
| Sonnet 5.5 | 0.5 | 0.1 | 0.2 | 0.01 |
| Haiku 4.5 | 0.25 | 0.05 | 0.1 | 0.005 |
| Measurement | Value | Provenance |
|---|---|---|
| Team Premium 5-hour window | 12,927,618 output-token equivalents; cache-hit rate 95.8% | 17,216,269 × 0.4 + 390,238,659 × 0.01 + 4,834 × 0.2 + 2,137,757, from the 2026-09-28 window. |
| Output only, no context | 12.9M Opus 5.5 or 5.2M Fable 5.1 output tokens per window | Derived from the window. |
| Team Premium week | 6.25 windows, 80.8M equivalents; Fable may use half | One window is 16% of the weekly limit. |
| Max 20x 5-hour window | 41.0M equivalents, 3.17 Team Premium windows; the page divides by exactly that | 2026-10-05, the only Max 20x measurement: railroad-review session b26f7f7c on the Max account, Opus 5.5 alone, 15,156,419 cache write, 353,337,991 cache read, 4,308 input, 1,876,747 output = 11,473,556 equivalents (88.75% of a Team Premium window), read 28% on the Max 5-hour meter. 11,473,556 / 0.28 = 41.0M. Anthropic's stated allowances (20 against 6.25) give 3.2. |
| Max 20x week | One window is 36% of the weekly limit, ~2.8 windows a week | The same session moved the Max weekly meter ~10 points: 10 / 28. Eyeballed; confirm with a reading before and after one session. |
| Team Premium check | Confirmed 2026-10-05 | Railroad-review session e82362c0 on the Team Premium account, Opus 5.5 alone, 10,143,570 equivalents: the page showed 78% of the window and 13% of the week, matching the Team Premium meters. |
The page offers only these two plans; Pro, Max 5x and Team Standard are unmeasured.
The plan meters each token class in proportion to its list price. One window fixes a total, not a per-class weight; a second window with a different mix would test it.