Eudaven Ads

Observability

Costs ?

Every LLM call logs its model, tokens, and dollar cost into pipeline_cost_events. The breakdown below shows where you're spending across the six pipeline nodes.

Total spend

$0.2583

LLM calls

18

Prompt tokens

47,351

Completion tokens

8,068

By node

Node Calls Cost Prompt tok Completion tok
brief_synth 2 $0.0720 12,704 2,262
synthetic_perf_scorer 3 $0.0602 7,475 2,521
brand_scorer 3 $0.0569 6,887 2,413
copy_polish 4 $0.0393 11,428 332
compliance_check 4 $0.0294 7,241 512
hook_diverger 2 $0.0005 1,616 28

Recent calls — last 18

When Node Model Cost Prompt Completion Brief
May 12 12:13:48 compliance_check anthropic/claude-sonnet-4-6 $0.0055 1,809 8
May 12 12:13:48 copy_polish anthropic/claude-sonnet-4-6 $0.0098 2,860 78 1c47885a…
May 12 04:11:05 synthetic_perf_scorer anthropic/claude-sonnet-4-6 $0.0203 2,492 852
May 12 04:10:50 brand_scorer anthropic/claude-sonnet-4-6 $0.0189 2,296 800
May 12 04:10:43 copy_polish anthropic/claude-sonnet-4-6 $0.0098 2,860 83 1c47885a…
May 12 04:10:43 compliance_check anthropic/claude-sonnet-4-6 $0.0077 1,811 150
May 11 13:37:49 synthetic_perf_scorer anthropic/claude-sonnet-4-6 $0.0199 2,489 828
May 11 13:37:35 brand_scorer anthropic/claude-sonnet-4-6 $0.0187 2,293 787
May 11 13:37:28 compliance_check anthropic/claude-sonnet-4-6 $0.0106 1,808 346
May 11 13:37:26 copy_polish anthropic/claude-sonnet-4-6 $0.0098 2,860 82 1c47885a…
May 11 13:37:24 hook_diverger nousresearch/hermes-4-70b $0.0002 795 14 1c47885a…
May 11 13:36:57 brief_synth anthropic/claude-sonnet-4-6 $0.0363 6,352 1,150 1c47885a…
May 11 02:52:50 synthetic_perf_scorer anthropic/claude-sonnet-4-6 $0.0201 2,494 841
May 11 02:52:35 brand_scorer anthropic/claude-sonnet-4-6 $0.0193 2,298 826
May 11 02:52:34 compliance_check anthropic/claude-sonnet-4-6 $0.0056 1,813 8
May 11 02:52:30 copy_polish anthropic/claude-sonnet-4-6 $0.0099 2,848 89 d7cc56a3…
May 11 02:52:29 hook_diverger nousresearch/hermes-4-70b $0.0003 821 14 d7cc56a3…
May 11 02:52:04 brief_synth anthropic/claude-sonnet-4-6 $0.0357 6,352 1,112 d7cc56a3…

How to use

Cost dashboard

Where the pipeline is spending money. Every LLM/image call writes a pipeline_cost_events row at emit-time. This page slices that table by node, day, and model.

Cost per accepted variant

brief_synth $0.05 + hooks $0.001 + polish $0.06 + compliance $0.003 + brand $0.005 + visual $0.34 ≈ $0.46

If a line item is way off, that node has a bug or got into a retry loop.

Node reference

  • brief_synthesizer — Sonnet, ~7K token KB context, $0.05/call
  • hook_diverger — Hermes 4 70B, 50 completions in parallel, $0.001/call
  • copy_polish — Opus, $0.06/call
  • compliance_judge — Sonnet, only fires for ambiguous rules
  • brand_scorer, synthetic_perf_scorer — Sonnet, rubric scoring
  • visual_direction + concept_render + final_render + vision_judge — image pipeline, see /admin for the daily budget knob

How to read the dashboard

  1. Top numbers are today's spend. Compare to yesterday and the 7-day average.
  2. The per-node bars show what's dominating cost — usually polish + visual.
  3. If you see a spike on vision_judge, someone is re-running concept-pick repeatedly.
  4. If copy_polish dominates, the compliance loop is iterating too much — check the iteration cap setting.

Throttling

Image-gen has a daily budget guardrail (editable at /admin/settings). Other nodes don't yet — if you need a hard cap, file it.