Hitting limits? Here is the arithmetic.
Subscription caps are a budgeting device: the plan meters you by session windows and weekly ceilings so the flat fee stays flat. Credits invert it: no reset schedule, your bill just tracks your work. The question is what your work costs, so we metered it.
One working week, two billing models
Subscription plan
Hard stop. The rest of the week waits for the reset.
API credits
No wall. The meter just adds up what you used.
Schematic, not a measured quota: where the wall lands varies by plan and week. That there is a wall is the design.
Provider limits change often and are rarely published precisely; treat any specific cap you read anywhere as dated. That volatility is what the alert list below is for.
One real agent session, metered
Numbers about caps are usually vibes. This is a bill: one Claude Code working session on API credits, five calls, priced by the provider meter.
Session ledger, replayed
Real bill- API call 1metered
- API call 2metered
- API call 3metered
- API call 4metered
- API call 5metered
Caching on, as the session really ran. The re-read discount is already in this number.
Each call bills tokens in and tokens out; the meter does not care what the call did. Call-by-call amounts are in the pricing writeup; this page keeps the sum.
Why caching moves this bill
An agent session reads far more than it writes: context, files, earlier steps, again on every call. This one read 78 tokens for every 1 it wrote. Caching discounts exactly that re-reading.
, the shape of agent work.
What a task costs when it is metered
One identical prompt, run the same day, billed on metered dashboards. Not a benchmark, a receipt: this is the spread a meter sees for the same job.
Second task, a game build, so you know the spread is not a one-off:
Meters bill verbosity too: on an identical task at identical rates, . Same prices, bigger bill.
Your week in credits
Weekly bill = tasks per day . Slide your workload.
- Fable 5 $10.23$0.1462 per task
- Opus 5 $8.11$0.1158 per task
- Opus 4.8 $5.07$0.0724 per task
- Sonnet 5 $2.87$0.0410 per task
- Pareto $1.37$0.0196 per task
- Haiku 4.5 $1.13$0.0162 per task
Your tasks are not this exact task; the shape is the point, not the penny. A heavy week costs more and finishes. A capped week just stops.
The rate card after the cap
Since July 7, 2026 the frontier model is credits-only, so this is the menu. Per million tokens, provider list.
| Model | Notes | ||
|---|---|---|---|
| Fable 5 | $10 | $50 | API only since Jul 7, 2026 |
| GPT-5.6 Sol | $5 | $30 | output 6x input |
| Opus 5 | $5 | $25 | launched Jul 24, 2026, same card as Opus 4.8 |
| Opus 4.8 | $5 | $25 | |
| Kimi K3 | $3 | $15 | cache-hit input $0.30; API live Jul 16, 2026 |
| Terra | $2.50 | $15 | |
| Sonnet 5 | $2 | $10 | until Aug 31; $3 / $15 from Sep 1 |
| Luna | $1 | $6 |
The cap timeline so far
Three moves, one direction. Hover or tab across the strip.
Fable 5 leaves Claude subscription plans. The frontier model becomes API-only: usage credits at $10 in and $50 out per Mtok. The cap conversation becomes a meter conversation.
The direction is one-way: the heaviest users get metered first. Planning around caps means re-planning every time they move.
Limits change without changelogs. Here is the log.
Kept by hand, dated, source-checked. The alert list gets each new entry before this page does.
SAMPLE ALERT Jul 7, 2026: Fable 5 left Claude subscription plans. API-only, usage credits, $10 in / $50 out per Mtok. What it does to the math: the heaviest tier now meters, so a heavy week is a bill, not a wall. Full record: unbiased.ai/limits/
One email when a major provider changes usage limits or plan terms: what changed and what it does to the math. Price moves have their own dated record and list on the AI price tracker.
One request in, one answer out: the blend picks per request, so you skip this choice entirely.
Get your credits matched