Sample dashboard built from stand-in CLIs (tests/fakes) by scripts/build_site.py. It shows the layout; the numbers, including “saving”, are not results. Back to ccx

Token and quota usage

Claude Code ⇄ Codex ⇄ DeepSeek via ccx · generated 2026-10-31 00:04 · last 30 days · amounts are API-equivalent USD (AE$), not your subscription bill

Total spend (AE$)
$0.0166
last 30 days · sub-agents only
Sub-agent spend
$0.0166
8 runs · cost/success $0.0028
Success rate
75%
7 judged by you or the lead · 0 probable (auto)
Lead cost per turn
–
median of 0 turns · ≈ – a turn · shorter conversations cost less (`ccx lead`)
Cache-read share
74%
of all input tokens
Codex credits
0.2
AE$ × 25 · quota stops 1
Estimated saving (model)
-$0.1504
95% CI -$0.1740 … -$0.1267 · n=6 · a model, not measured; measured totals so far about even (docs/measured.md). The model: the task-sized part of the sub-agent's tokens (fresh input + output, minus its start-up) at the lead's price; the old model (all its tokens at the lead's price, start-up included) gave -$0.1382. Measured (H11): what the lead saves is the requests it doesn't make — about one per small task, only when it trusts ccx's check. Measure yours with `ccx trial start`

Spend per day, by who spent it

Claude · leadClaude · sub-agentCodex · leadCodex · sub-agent
$0.0166$0.0125$0.0083$0.0042$010-31
Table view
day Claude · leadClaude · sub-agentCodex · leadCodex · sub-agent
2026-10-31$0$0.0094$0$0.0073

Lead-side usage not read (run with transcripts available to include it).

Tokens by type

fresh inputcache writecache readoutput
Claude · sub-agent12,330Codex · sub-agent5,400
Table view
stream fresh inputcache writecache readoutputAE$
Claude · sub-agent302,7009,000600$0.0094
Codex · sub-agent1,60003,200600$0.0073

Cache reads bill at about 0.1× input; Claude's headless start-up is written to a 1-hour cache (2× input) on first use, so delegations close together are cheaper.

Cost per task: built-in sub-agents vs ccx sub-agents

who did the sub-tasktasksmean AE$median AE$
ccx → claude3$0.0031$0.0046
ccx → codex5$0.0014$0.0024

Built-in sub-agents spend the lead's own quota; ccx sub-agents spend the other tool's quota. Compare like with like: similar tasks, same period.

Where it was wrong — this week vs last week

evaluation runs left out (`ccx report --include-evals` to count them)

By tier and kind of work

agenttierkindjudgedwrongreworklast wk judgedwrongrework
claudehaiku:lowcode1100%100%–––
claudesonnet:mediumcode20%0%–––
codexgpt-6-luna:lowcode10%0%–––
codexgpt-6.1-sol:highcode20%0%–––
codexgpt-6.1-sol:mediumcode10%0%–––

Routing prediction

Brier score (0 is perfect, 0.25 is a coin flip): this week: – (n=0) · last week: – (n=0). Right = judged right and done by the rung routing picked.

Probable verdicts overturned later

this week: 0/0 (–) · last week: 0/0 (–)

Delegated runs

daytaskwho model → escalationresulttokensAE$est. saving (model)
2026-10-3120261031-000416-7729claude → codexgpt-6.1-sol:mediumerror · quota0$0–
2026-10-3120261031-000416-322ccodex → claudesonnet:medium✓ verified4,110$0.0046-$0.0175
2026-10-3120261031-000415-5c1bclaude → codexgpt-6.1-sol:high✓ verified1,350$0.0024-$0.0294
2026-10-3120261031-000415-47c2codex → claudehaiku:low✗ rejected4,110$0.0002–
2026-10-3120261031-000414-f16fclaude → codexgpt-6.1-sol:medium✓ verified1,350$0.0024-$0.0294
2026-10-3120261031-000414-7baaclaude → codexgpt-6.1-sol:medium → gpt-6.1-sol:high✓ verified1,350$0.0024-$0.0294
2026-10-3120261031-000414-5c8bcodex → claudesonnet:medium✓ verified4,110$0.0046-$0.0175
2026-10-3120261031-000413-1f6cclaude → codexgpt-6-luna:low✓ verified1,350$0.0001-$0.0271

Estimated saving (a model, not measured) = the task-sized tokens (fresh input + output, minus the sub-agent's start-up: ~20k fresh tokens for Codex) at the lead's model price − sub-agent cost − brief/review overhead. Negative means delegating cost more than doing it yourself would have. Measured totals so far were about even (docs/measured.md).