← Control Room

Dispatch · Internal · 💲

Compute & Budget.

The whole money picture up top — what we're spending and the headroom — then the services we're considering, then the compute detail underneath. Numbers pulled from the compute ledger + the live account inventory; nothing here is fabricated.

Figures as-of: monthly spend live from accounts.config.json · weekly usage 2026-07-07 · per-task tokens 2026-07-09 · a snapshot, not a live meter (see the honesty box)


01 Total monthly spend — the one financial view

Everything in one place: fixed subscriptions + variable/usage-based API = the real monthly total. ✅ = confirmed · ~ = estimate. Fixed rows are live-sourced from the account-health inventory; variable rows are estimated from the compute ledger + connector runsheets.

A · Fixed subscriptions — flat, bills whether used or not
ServiceWhat$/mo
Claude / Anthropic (Max)Dispatch's backbone — every session + scheduled task~$200 *
Gemini / Google AI PlusJustin's sub (already paid)$8 †
DigitalOcean VPS24/7 host — Hermes + Umami~$6 *
Domain — fast2future.comregistrar, amortized~$1
GitHub · Cloudflare · Buttondown · Gumroad ×2 · Seedance · socialsfree tiers$0 ✅
Fixed subtotal≈ 93% is Claude
B · Variable / usage-based API — bills per call; today near $0
ServiceStatus$/mo nowif active (est.)
Claude APIruns on the Max sub — no separate metered API$0 ✅
OpenRouter (Hermes)pay-per-token; capped $0, Hermes not live$0 ✅~$5–20 (capped)
Hunter.io + Tomba/Reoon/ZeroBouncelead verify/enrich — free tiers (~800 verify/mo free)$0 ✅$0, or $34+ paid
ElevenLabsTTS — free tier; $5 Starter to publish audio commercially$0 ✅~$5
Exa (neural search)~$10 free credits, then pay-as-you-go$0 ✅usage
fal.ainot wired — no active account in the repo$0usage if added
Variable subtotal (today)armed but idle
~$215/mo total real monthly spend (fixed ~$215 + variable ~$0)

~93% is Claude. The variable/usage layer is armed but idle — OpenRouter capped at $0, and enrichment + TTS + search all sit on free tiers. So today's total ≈ the fixed subscriptions. Headroom: Claude's weekly capacity is 76% used (full gauge in §03 below); the one $ line to watch is OpenRouter, the moment Hermes goes live (capped, never silent).

* Confirm: Claude Max tier ($100 vs $200) + DO droplet price (best estimates). † Gemini $8 → ~$5 at next renewal (Google AI Plus cut to $4.99, eff. 2026-06-08). Fixed rows live-sourced from accounts.config.jsonaccount_health.py; variable estimates from the compute ledger + CONNECTOR_RUNSHEETS.md — reconcile against real balances (openrouter_cost.py) when any usage-based service goes live. This section is the single source of truth for "what are we spending per month, total."

02 Considering — services we're exploring (reallocate, don't stack)

Firsthand exposure to other frontier ecosystems, kept ≈ flat. Prices verified July 2026 (they move fast — re-check at signup). Nothing is bought without your tap (money = hard-gate).

The honest reallocation reality

  • There's almost nothing to reallocate from — everything but Claude is free or tiny. So "flat" means one of: (a) wire pay-as-you-go APIs into the engine (pennies/mo, not $20 subs — and it extends Dispatch), (b) rotate one consumer sub at a time (cancel before the next), or (c) confirm/trim the Claude tier.
  • 💡The big unlock: via the repo's existing LiteLLM door (loop/llm_client.py), any API key drops in as a one-line completion(model=…) — same pattern as the OpenRouter key. Multi-LLM through the engine costs ~$0–10/mo metered, vs. ~$60–90/mo stacking consumer subscriptions.
🔌 = plugs into Dispatch (API via LiteLLM) · 🖥️ = CLI runs on the machine (alongside Claude Code) · 📱 = standalone consumer app (your personal use, no engine hook)
EcosystemReal priceWhat it teaches / builds vs ClaudeEngine fit
Claude (keep)~$100–200/moThe backbone — non-negotiable. Everything else is additive.🔌 is the engine
Grok (xAI)API $2/$6 per M · SuperGrok $30/moReal-time X/social-grounded reasoning; "Heavy" multi-agent mode. Up to $175/mo free dev credits.🔌 API (LiteLLM xai/) · 📱 app
Nous Hermes~$0.13–3 per M (OpenRouter)Open-weight, steerable/less-restricted agentic behavior. ⚠️ SEPARATE from your internal "Hermes."🔌 API (OpenRouter — already wired)
OpenAI / ChatGPTPlus $20 · Pro $100–200 · API sep.Sora (video), Operator (browser agent), Codex, Workspace Agents — a different frontier lineage.🔌 core API · 🖥️ Codex CLI · 📱 Sora/Operator/voice
Meta / Muse SparkConsumer FREE · Muse Spark API $1.25/$4.25 per MCheap agentic computer-use (the FB-Marketplace-listing demo). "Meta One" $8/$20 = app-only.🔌 Muse Spark API (US preview) · 📱 Meta AI app
Gemini / Google AI already paid$8 → ~$5/mo (AI Plus)Get more from what you own: Gems, AI Studio (free Flash), Veo/Nano-Banana + Antigravity CLI.🔌 Gemini API · 🖥️ Antigravity CLI · 📱 app/Gems
Abacus.AI (ChatLLM/DeepAgent)$10–20/mo · enterprise API ~$5K/mo"Personal superintelligence" — one app over many models + DeepAgent (build custom agents) + CodeLLM vibe-coding. Credit-based (⚠️ 75% usage-lock rule).📱 app (wireable API is enterprise-tier only)
DeepSeekAPI ~$0.14/$0.28 per M · 5M freeNear-free frontier reasoning → learn "run 1,000 experiments for pennies" (bulk/agentic loops).🔌 API (LiteLLM, one-line)
Perplexity Pro$20/mo · Sonar APILive cited search as the default unit — grounded/RAG answers without building retrieval.🔌 Sonar API (LiteLLM)
MistralLe Chat $15/mo · API sep.EU/open-weight lineage + Codestral (code model to benchmark vs Claude); native MCP agents.🔌 API · MCP connector

⚠️ Worth flagging

"Grok 4.5 with Hermes": Grok = xAI. "Hermes" almost certainly = Nous Research's open Hermes model (a separate thing) — NOT your internal Dispatch "Hermes" agent. They connect, though: your Hermes framework could run on Grok or Nous-Hermes as its backing model. Confirm which you meant.

"anti gravity cli" = Google Antigravity CLI — a real, locally-run agentic CLI (successor to the now-retired Gemini CLI), free tier + ~$20 Pro. Runs alongside Claude Code; it's a CLI, not an API you wire in.

Abacus.AI at $10–20/mo is a polished consumer app that aggregates models you already have (GPT/Claude/Gemini) + an agent-builder (DeepAgent). Its real developer API is enterprise-tier (~$5K/mo), so at your price point it doesn't cheaply extend Dispatch — great for hands-on tinkering, not an engine hook. Watch the credit-based "75% usage-lock" rule.

📌 The flat-budget pick

Best learning-per-dollar that EXTENDS Dispatch: wire 2–3 pay-as-you-go API keys into llm_client.pyDeepSeek (near-free bulk), xAI Grok (free dev credits), Nous Hermes (OpenRouter, already wired). ≈ $0–10/mo metered, no new subscription, each becomes a skill/loop.

For hands-on consumer tinkering (Operator, Sora, Antigravity CLI, Abacus DeepAgent, the Grok app): rotate ONE ~$10–20 sub at a time. The Gemini drop to ~$5 frees a little headroom.

No subscriptions created — this is a menu for your tap. Name one and I'll stage the wiring (API key → skill/loop) or the signup steps.

03 Weekly budget & headroom

On the flat-rate Claude Max plan there is no per-token dollar meter. The weekly "all models" usage bar IS the budget signal.

76% of the weekly "all models" cap used
0%ceiling → 85%100%

76% used as of 2026-07-07 (Justin-reported from Claude Max → Settings → Usage — there is no API to read this, so it is updated by hand in compute-usage-cap.json). The dashed line is a suggested 85% ceiling to keep a safety margin — no ceiling is set yet; tell me your real number and I'll track against it. A separate weekly "Sonnet-only" bucket exists and draws independently. The Fable window closed ~2026-07-12.

$0 variable
Flat Max subscription — no metered per-token cost.basis: plan type
~1.20B
Scheduled-task tokens (with cache), measurement window 06-22 → 07-09.MEASURED · floor
~4.43B
Interactive-session tokens (Justin's direct work) — 93% cache-read.MEASURED · floor

04 Model routing — rule vs reality

Compute discipline says default Sonnet, Opus only on explicit override. Here's what actually fires.

The rule

Default = Sonnet 🔵. Opus 🟣 requires an explicit override + a logged reason. Haiku ⚪ for scan/classify. Every unnamed task should be Sonnet.

Enforced by COMPUTE_ENFORCEMENT.md (pre-spawn gate) + the weekly weekly-compute-audit that reads the ledger.

The reality 🔴

Scheduled tasks ≈ 100% Opus 4.8. Verified across 9+ post-pin fires (07-08 / 07-09): every one ran Opus. In-prompt model pins are ignored by the runtime — native tasks inherit the app-session model at fire time.

The only real lever: set the Claude app's session/default model to Sonnet before scheduled tasks fire (UI-only — can't be set from a file).

Whole-machine model mix — by API-call count, 07-09

Opus 4.879%
Sonnet 4.69%
Fable 56.5%
Sonnet 55%
Haiku 4.50.06%

⚠️ This is call-count share, not token-cost share — a per-model token-cost breakdown is not yet computed (instrumentation gap).

05 Top cost centers

Scheduled tasks by total tokens moved (input + output + cache), MEASURED from local transcripts, 2026-07-09.

Top scheduled tasks · window 2026-06-22 → 07-09 · total incl. cache-read (billed at a steep discount)
TaskRunsTotal tokensAvg / runModel
assess-advance-overnight113468.6M4.1Mopus 🟣
assess-advance-heartbeat76246.7M3.2Mopus 🟣
the-forerunner25130.7M5.2Mopus 🟣
automation-audit-loop1675.7M4.7Mopus 🟣
assess-advance-backstop6464.4M1.0Mopus 🟣
best-of-best-radar1862.6M3.5Mopus 🟣
agent-infra-watchlist1858.3M3.2Mopus 🟣
body-watchlist1756.8M3.3Mopus 🟣
morning-brief1951.6M2.7Mopus 🟣
assess-advance trio253opus 🟣

The assess-advance trio (overnight + heartbeat + backstop) is the single largest center — ~65% of all scheduled compute. Cadence was cut on 07-07 (overnight 28 → 11.5 fires/24h, ~38% fleet-wide reduction) — the run counts above cover the full window, so post-cut burn is lower than the averages imply.

06 What's real vs. what needs building

The honesty box — so no number here is trusted further than it earns.

✅ Measured & trustworthy

  • Per-task token totals — parsed from real API usage fields in local transcripts. Exact, re-runnable (compute_usage_audit.py).
  • Model per fire — read from each transcript's message.model. This is how we know pins are ignored.
  • Monthly subscription costs — from the live account inventory (accounts.config.json); free tiers verified $0.
  • Weekly usage % — real, but hand-entered from the Settings→Usage screen (no API exists).

⚠️ Floors / estimates — would need instrumentation to show exactly

  • Real dollar compute spend: none — the Max plan is flat-rate with no per-token meter. The weekly % is the only budget proxy. A true $ figure would need API billing (different plan) or Anthropic exposing it.
  • Variable-API spend: ~$0 today is real, but the "if active" column is estimate — reconcile against real balances (openrouter_cost.py) when anything goes live.
  • Subagent tokens: not attributed to parent tasks (~978 subagent transcripts). Every token total is a floor; heavy-delegation tasks cost materially more.
  • Continuous trend: these are dated snapshots (07-07, 07-09), not a time series. A real burn-rate chart needs the audit to log a periodic snapshot to a history file — not yet built.
  • Live auto-refresh: compute figures are hand-refreshed; the monthly-cost total already auto-computes from the inventory. Full auto needs compute_usage_audit.py to emit JSON this page reads — a clean next step if you want it.