# RemakeBench Subscription Capacity Index — Alibaba Token Plan Personal Lite v1

Status: **Provisional measurement**

Evidence cutoff: **2026-07-21T16:56:02Z**

Record: **Alibaba Cloud Model Studio · Token Plan Personal Edition Lite · qwen3.8-max-preview · Qwen Code 0.20.0 · Singapore**

## What is official

At the cutoff, Alibaba's official Personal Edition page published a limited-time Lite price of $6/month, 700 Credits in a five-hour sliding window and 2,500 Credits in a seven-day sliding window. Consumption older than the window is released; reaching either quota pauses service until capacity is released. The dated official-source projection is included in this bundle.

The current page did not publish the older non-promotional price or a Concurrent Agents value, so both are unavailable here. Live official evidence wins over stale local wording.

## What is measured

The workload records are privacy-safe projections of five pinned local receipts created before the cutoff. Each source receipt is represented by a SHA-256 digest; raw prompts, completions, reasoning, session identifiers, provider exports, credentials and local paths are excluded.

Output-token fields include the provider-reported thoughts subset. Thoughts are not added a second time when total processed tokens are calculated.

- Jungle Temple: weekly meter 12% → 16%; 100 nominal Credits; 75–125 Credit endpoint interval. The run crossed the five-hour reset, so no five-hour delta is published.
- Space Flight attempt 1: after-only five-hour meter 14%; 98 nominal Credits; conservative 91–105 Credit range. The exact 4,438,918-token local ledger matched the provider's rounded five-hour token display, but no explicit zero-percent pre-run capture exists.
- MacBook clean attempt: five-hour meter 1% → 10%; 63 nominal Credits; reconciled 56–70 Credit range. The attempt failed at HTTP 413 and assisted recovery also failed.
- Four-activity aggregate: five-hour 3% → 42% and weekly 1% → 12%; reconciled midpoint 274 Credits, range 266–280. It spans four activities and is not allocated to individual tasks.
- Space Flight attempt 2: 4,277,142 processed tokens; HTTP 413 after hardware visual iteration. No post-run Credit meter exists, so Credit use and same-mix capacity are unavailable.

## Credit arithmetic

A whole-percent reading has ±0.5 percentage-point endpoint uncertainty. A delta between two whole-percent endpoints therefore has ±1 percentage point. Credits are calculated with integer arithmetic from the official allowance and the displayed delta. Where five-hour and weekly receipts describe the same interval, their Credit ranges are intersected.

Same-mix capacity divides the official 700- or 2,500-Credit allowance by the measured workload Credit point. The low repetition bound uses the high Credit bound; the high repetition bound uses the low Credit bound. Values are rounded to one decimal using integer tenths. These are workload-specific linear estimates—not promised builds, reliability claims or universal token limits.

## Monetary valuation withheld

The records preserve fresh input, cached input and output separately. Alibaba did not publish a first-party qwen3.8 rate card that values all three categories consistently, including cached input. A third-party scenario found in lower-level receipts is not used or extrapolated. The public monetary estimate therefore has status `withheld`, reason `first_party_qwen38_cache_rate_unavailable`, and null numerical fields. Missing cached-input pricing never falls back to ordinary input pricing.

## Promotions and quality

The qwen3.8 preview and night Credit discounts are recorded as mutable conditions, separate from the base allowance. Credits can vary with model, token category, thinking/tool behavior, time-sensitive promotions and tool calls. Price does not affect RemakeBench's quality score. Failed attempts consume capacity but cheap failure is not value.

## Evidence and privacy

Six first-party meter screenshots were decoded and losslessly encoded as PNG with Sharp 0.34.5. Every ancillary PNG chunk was removed; only IHDR, IDAT and IEND remain. Content-addressed filenames, byte lengths, dimensions, source hashes and decoded-pixel hashes are recorded. The images were reviewed for account identity, credentials, billing identifiers, private URLs and local paths.

The evidence cutoff excludes any active or incomplete Qwen activity after 2026-07-21T16:56:02Z. Published v1 files and artifacts are immutable; corrections require v2 or later.