tokencur observatory

What the maintainer actually pays for AI subscriptions vs what the same usage would cost at API list prices — real fees, showback valuation and counterfactuals, kept apart. Aggregates only: no workspaces, no sessions, no content.

History gap · Claude Code — no usage on record before 2026-09-24. Claude Code deletes session transcripts after 30 days, and this history was gone before tokencur kept a ledger. The 2026-07-21 snapshot, which still had it, recorded $750.56 of Claude Code usage value (2026-06-04 to 2026-07-21). Its subscription fee is not counted across the gap, so lost records don't read as months paid for nothing.

Correction · 2026-10-02 — Subscription fees were counted across the whole window, as if all three plans had been paid every day of it. Each plan now counts only while it was active: Codex CLI through April 2026, Kimi Code from June to September 2026, and Claude Code from 2026-09-24, upgraded from $20 to $100 a month on 2026-10-02. For 2026-02-07 to 2026-10-02 the page had published $240.49 of estimated outlay and 3.5× subscription leverage; counted plan by plan, the same window comes to $96.59 and 8.7×.

Correction · 2026-10-02 — Codex usage was counted about twice: Codex re-sends each usage report, and earlier snapshots counted every copy. Codex usage value went from $685.66 to $332.95; every Codex session now reconciles with Codex's own running total. Details in the changelog.

What is actually paid

$100.00
subscriptions / month now
$99.88
est. outlay over 7.85 months
9.5×
subscription leverage
89.5%
effective discount vs API

Subscription fees really paid, each counted only while its plan was active (Claude Code $20/mo from 2026-09-24 to 2026-10-01, then $100/mo from 2026-10-02; Codex CLI $20/mo until 2026-04-30; Kimi Code $10/mo from 2026-06-21 to 2026-09-30): actual money. Leverage = API-equivalent usage value ÷ estimated outlay over the same window (fees not counted across the history gap above).
Claude Code: Pro for about a week before the upgrade, counted from 2026-09-24, the first day on record.
Codex CLI: ChatGPT Plus, no longer paid. Its cancellation date was not recorded, so it counts through April 2026, the month Codex CLI was last used.
Kimi Code: No longer paid. Its dates were not recorded, so it counts from the first day Kimi Code was used through September 2026, the month it was last used.

$2.69
billed by providers
20
billed charges

Provider bills (RunPod Pods $2.69) are real money too, for compute rather than coding-agent usage, so they stay out of subscription leverage.

Usage value — showback, not money spent

$950.17
API-equivalent usage value
53
active days
$17.93
avg value / active day
11,579
assistant messages

Subscriptions bill a flat fee, not per token. These figures value the usage at API list prices: what this work would have cost on pay-per-token billing.

Counterfactuals — at API list prices

$6,704.73
avoided by provider caching
$213.81
right-sizing headroom (what-if)

Cost avoided by caching is a property of the provider's architecture, not an action taken. Under a flat subscription, right-sizing buys rate-limit headroom, not dollars — it would save real money only on API billing.

Daily usage value by service

Claude CodeCodex CLIKimi Code
$0$50$100$150$20002-0702-1702-2703-1008-0610-0110-03
table
dayClaude CodeCodex CLIKimi Code
2026-02-07$0.00$2.00$0.00
2026-02-08$0.00$1.37$0.00
2026-02-09$0.00$2.48$0.00
2026-02-10$0.00$8.33$0.00
2026-02-11$0.00$5.29$0.00
2026-02-12$0.00$12.17$0.00
2026-02-13$0.00$2.52$0.00
2026-02-14$0.00$7.71$0.00
2026-02-15$0.00$13.94$0.00
2026-02-16$0.00$4.75$0.00
2026-02-17$0.00$0.79$0.00
2026-02-18$0.00$16.26$0.00
2026-02-19$0.00$1.47$0.00
2026-02-20$0.00$31.64$0.00
2026-02-21$0.00$10.31$0.00
2026-02-22$0.00$0.60$0.00
2026-02-23$0.00$1.36$0.00
2026-02-24$0.00$0.77$0.00
2026-02-25$0.00$0.52$0.00
2026-02-26$0.00$0.15$0.00
2026-02-27$0.00$5.79$0.00
2026-03-01$0.00$4.93$0.00
2026-03-02$0.00$1.63$0.00
2026-03-03$0.00$0.70$0.00
2026-03-04$0.00$0.80$0.00
2026-03-05$0.00$3.68$0.00
2026-03-06$0.00$28.97$0.00
2026-03-07$0.00$13.84$0.00
2026-03-08$0.00$36.81$0.00
2026-03-09$0.00$24.56$0.00
2026-03-10$0.00$79.31$0.00
2026-03-11$0.00$6.94$0.00
2026-03-16$0.00$0.19$0.00
2026-03-17$0.00$0.03$0.00
2026-04-25$0.00$0.34$0.00
2026-06-21$0.00$0.00$0.17
2026-06-23$0.00$0.00$1.20
2026-06-24$0.00$0.00$0.31
2026-06-28$0.00$0.00$0.42
2026-06-29$0.00$0.00$0.11
2026-08-06$0.00$0.00$0.11
2026-08-10$0.00$0.00$0.11
2026-09-04$0.00$0.00$0.01
2026-09-24$8.91$0.00$0.00
2026-09-25$5.69$0.00$0.00
2026-09-26$21.14$0.00$0.00
2026-09-27$71.88$0.00$0.00
2026-09-28$106.04$0.00$0.00
2026-09-29$26.14$0.00$0.00
2026-09-30$79.85$0.00$0.00
2026-10-01$95.62$0.00$0.00
2026-10-02$189.02$0.00$0.00
2026-10-03$10.49$0.00$0.00

Usage value by model

table
modelvalue USDtokens
claude-opus-5-5$603.711,474,504,594
gpt-5.4$189.95341,017,638
gpt-5.3-codex$139.71405,185,814
claude-fable-5-1$7.361,982,985
gpt-5.2-codex$2.495,414,267
claude-opus-5$2.46264,415
moonshot-ai/kimi-k2.7-code-highspeed$2.449,925,718
claude-opus-4-8$1.26686,048
gpt-5.1-codex-mini$0.8115,048,582

Where the value concentrates (token type)

Recommendations

kindrecommendationUSD (list prices)% of baseline
achievedPrompt caching on claude-opus-5-5$5,391.2791.8%
achievedPrompt caching on gpt-5.4$685.1990.0%
achievedPrompt caching on gpt-5.3-codex$593.8490.0%
potentialRight-size claude-opus-5-5 → claude-sonnet-5-5$157.4326.1%
potentialRight-size gpt-5.4 → gpt-5.3-codex$50.6626.7%
achievedPrompt caching on claude-fable-5-1$15.0378.4%
achievedPrompt caching on gpt-5.2-codex$7.8790.0%
achievedPrompt caching on moonshot-ai/kimi-k2.7-code-highspeed$7.2683.2%
potentialRight-size claude-fable-5-1 → claude-opus-5-5$4.2457.6%
achievedPrompt caching on gpt-5.1-codex-mini$2.9890.0%
achievedPrompt caching on claude-opus-4-8$2.4171.5%
potentialRight-size claude-opus-5 → claude-sonnet-5$1.4860.0%
achievedPrompt caching on claude-opus-5$-1.12-85.1%