RU

How much AI coding fits in a subscription

A subscription shows a percentage bar. I converted it into dollars: took tokens from local Claude Code and Codex logs and multiplied them by public API prices. The result: what one percent of the week is worth, where it goes and which model is the better deal.

1% of a Max 20x week

at API list price

Whole week

on a $200/month plan

Requests > 300k

of all spend — long context

Auto-compact at 267k

spend, simulated over 9 days

Data as of Oct 11, 2026 · Claude Max 20x · Codex Pro Lite / Plus

What one percent of the bar is worth

List-price dollars per 1% of the weekly limit: tokens burned before the bar peaked, divided by the peak percent. A Claude percent is worth 5–10× a Codex one at a similar subscription price.

List $ per 1% of the week
= $4.0small = remainder

Windows differ in length: the limit was sometimes reset early. Codex August is understated — the app purged part of the logs. Windows where the bar stayed below 60% are dropped as too noisy.

Table
WindowPlanPeak$ to peak$ per 1%
Sep 11–Sep 18Claude Max 5x100%$1,837$18.37
Sep 18–Sep 23Claude Max 5x100%$2,007$20.06
Sep 23–Sep 25Claude Max 20x91%$2,460$27.03
Sep 25–Sep 28Claude Max 20x100%$2,612$26.12
Sep 28–Oct 2Claude Max 20x86%$2,033$23.64
Oct 2–Oct 9Claude Max 20x94%$2,207$23.47
Oct 9–Oct 11Claude Max 20x20%$513$25.64
Aug 1Codex Pro Lite85%$148$1.74
Aug 9Codex Pro Lite100%$640$6.39
Aug 11Codex Pro Lite66%$276$4.18
Aug 13Codex Pro Lite100%$276$2.76
Aug 20Codex Pro Lite73%$176$2.41
Aug 31Codex Plus100%$29$0.29
Sep 9Codex Pro Lite100%$370$3.70
Sep 12Codex Pro Lite100%$368$3.68
Sep 14Codex Pro Lite100%$334$3.34
Sep 16Codex Pro Lite100%$393$3.93
List dollars per $1 of subscription per month
= ×3small = remainder
List dollars per $1 of subscription per month
CategoryValue
Claude Max 5x · $100/mo · week ≈ $1,922×84
Claude Max 20x · $200/mo · week ≈ $2,488×54
Codex Pro Lite · $100/mo · week ≈ $369×16
Codex Plus · $20/mo · week ≈ $29×6

Median of weeks where the bar nearly filled, × 30.4 / 7, divided by the price. On 5x I ran Opus 5, on 20x Opus 5.5: different models, so 5x and 20x are not directly comparable.

The week, day by day

Oct 2–Oct 10: $2,732 at list price. Lasting until the reset takes an even pace of $335 a day. Three days (Oct 5, Oct 7, Oct 10) ate 63% of the spend.

List $ per day, Claude
= $100small = remainder
Table
Day$% of weekRequests
Oct 2$271.1%422
Oct 3$1998.5%2,443
Oct 4$25110.7%2,659
Oct 5$76232.5%8,890
Oct 6$1446.1%1,813
Oct 7$45119.2%5,685
Oct 8$1958.3%3,642
Oct 9$2048.7%2,662
Oct 10$49921.3%7,507

Long context is the main cost

Every request re-reads the whole accumulated context, so a 300k-token conversation is expensive even with caching. Compacting at 267k would have cut 9 days of spend by 10%.

$ by prompt size, since Oct 2
= $40
$ by prompt size, since Oct 2
CategoryMain chatSubagents
<50k · 1%$0$19
50–100k · 9%$31$212
100–200k · 27%$283$444
200–300k · 28%$358$409
>300k · 36%$551$430
Compact threshold 267k

≈ −10% · +251 compactions in 9 days

Compact threshold 350k

≈ −8% · +119 compactions in 9 days

After compaction

median context

The cache goes cold while an agent waits

A subagent's cache lives 5 minutes, the main chat's an hour. Idle longer and the whole context is written again at write price. Since Oct 9 that cost $96: subagents $78, main chat $19.

What a window’s bill is made of
= $100
Table
WindowTotalReadWriteRebuildOutput
Sep 11–Sep 18$1,860$1,229$284$83$264
Sep 18–Sep 23$2,018$1,415$237$99$266
Sep 23–Sep 25$2,497$1,261$519$133$584
Sep 25–Sep 28$2,628$1,330$572$170$557
Sep 28–Oct 2$2,108$1,283$458$131$235
Oct 2–Oct 9$2,220$1,221$562$99$339
Oct 9–Oct 11$516$250$87$96$84
Why subagents went idle, Oct 9–11
= 1 %
Why subagents went idle, Oct 9–11
PartValueShare
Waiting on Bash or a background task · 73 times$65.785%
Messaged an agent after a pause · 14 times$1114%
Other · 3 times$0.91.2%

The fix is simple: Bash waits at most 4 minutes and longer work goes to the background; an agent that returned long ago is not woken up — a fresh one gets a short brief.

The price of one subagent

A heavy Opus engineer costs as much as a hundred Haiku runs: long sessions, context near 250k. 86% of subagent turns make a single tool call, 66% in the main chat — and every extra turn re-reads the whole context.

Average list $ per agent, since Oct 2
= $0.50small = remainder
Average list $ per agent, since Oct 2
CategoryValue
opus-engineer · 17 runs · median 102 req · peak 248k$14.74
general-purpose · 134 runs · median 45 req · peak 209k$3.55
workflow-subagent · 275 runs · median 19 req · peak 142k$2.51
opus-verifier · 24 runs · median 45 req · peak 137k$2.44
Explore · 25 runs · median 20 req · peak 136k$1.45
haiku-implementer · 13 runs · median 72 req · peak 175k$0.18
haiku-finder · 2 runs · median 32 req · peak 173k$0.05

My own model index

The Artificial Analysis index is half science and trivia. I care about agentic coding, understanding the task and long-running work, so I re-weighted the same benchmarks: agentic coding 45%, understanding and long work 45%, not hallucinating 10%. Science is dropped.

Index vs cost per task
  • Sonnet 5.5
  • Opus 5.5
40 50 60 70 $0$1$2$3$4 $ per AA index task high · 58 xhigh · 65 high · 49 xhigh · 62

In points per $1, Sonnet xhigh and Opus high are level (31 vs 32) — switching to Sonnet saves nothing. Sonnet high is the cheapest but only Haiku-max quality. Opus xhigh is 7 points smarter than high at almost twice the price. Hollow dot is an estimate.

Official index → mine
  • Sonnet 5.5
  • Opus 5.5
Artificial Analysis my weights 585654 67 Opus 5.5 max 69 Sonnet 5.5 max 65 Opus 5.5 xhigh 58 Opus 5.5 high

Without science, Sonnet overtakes Opus. Opus medium (≈52) is clearly worse than high — saving on Opus effort does not pay.

All effort levels
ModelEffortMy indexAgenticUnderstanding$ per AA taskPoints per $1
Sonnet 5.5max696679——
Opus 5.5max676280——
Opus 5.5xhigh656176$3.4619
Sonnet 5.5xhigh626069$2.0131
Opus 5.5high585964$1.8232
Opus 5.5medium~52—51——
Sonnet 5.5high~495249$0.88~56
Haiku 5.5max~433450——

Experiment: what the new rules change

On Oct 11 I turned on the rules: Bash waits up to 4 minutes, context auto-compaction, subagents matched to the job (search on Haiku, tricky logic on Opus). The clean measurement week is Oct 16–23, compared with Oct 2–9.

Metric Before · Oct 2–9 After · Oct 16–23
$ per 1% of the week $23.5 measuring
$ per day $317 measuring
Share of $ in > 300k requests 36% measuring
Rebuilds after idle $99 measuring
opus-engineer, $ per agent $14.74 measuring

Method and caveats

  • Dollars are public API prices applied to local tokens, not the vendors’ internal meter. Cache and reasoning may be weighted differently there.
  • claude.ai chat is not in local logs — browser usage makes a percent look cheaper.
  • The limit bar is sampled every 15 minutes, with occasional day-long gaps.
  • Opus 5.5 per 1M tokens: input $4, cache read $0.20, write $5 (5 min) / $8 (1 hour), output $20.
  • Some high/xhigh benchmark values come from screenshots; “~” marks estimates. Elo is converted to win rate against 1600.