FasClaw Economics • March 2026

TOKEN &
COST ANALYSIS

Per-run token estimates, API cost breakdown by model, runtime benchmarks, and subscription plan comparison — so you know exactly when API beats Pro/Max.

7 Workflows
5 Models
3 Plans
10 Bug Vectors
01

Model Pricing Reference — Per 1M Tokens

Anthropic
Haiku 4.5
Input
$1.00
Output: $5.00 / 1M
FasClaw heartbeat model
4–5× faster than Sonnet
★ Default
Sonnet 4.6
Input
$3.00
Output: $15.00 / 1M
FasClaw primary model
Best quality/cost ratio
Anthropic
Opus 4.6
Input
$5.00
Output: $25.00 / 1M
Most capable
Complex reasoning tasks
OpenAI
GPT-4o mini
Input
$0.15
Output: $0.60 / 1M
Cheapest option
Good for simple queries
OpenAI
GPT-4o
Input
$2.50
Output: $10.00 / 1M
Comparable to Sonnet
Slightly cheaper input
Batch API discount: All models offer 50% off via Batch API (async, not real-time). Prompt caching saves up to 90% on repeated system prompts — highly recommended for FasClaw's base system prompt (~4,500 tokens repeated on every call).
02

Token Estimate + Cost per Run

Base overhead per call (every workflow): OpenClaw system prompt (~1,500 tok) + Tool definitions (~3,000 tok) = ~4,500 tokens of fixed input per run. With prompt caching, this drops to ~450 tok effective cost. All estimates below assume uncached (worst case). Multi-turn = tool calls stack context across turns.
Workflow Model Input Tokens Output Tokens Turns Haiku 4.5 Sonnet 4.6 GPT-4o mini GPT-4o Opus 4.6
⏱ Heartbeat Check
Haiku health ping every 30min
Haiku 2,000 80 1 $0.0004 $0.0072 $0.0003 $0.0056 $0.012
📋 Telegram Query
"สถานะ?" / "ลูกค้าใหม่?" / simple lookups
Sonnet 8,500 400 2 $0.011 $0.032 $0.0015 $0.025 $0.053
📌🤖 Calendar Worker
Scan events → Context Router → Lead DB → Create Task
Sonnet 12,000 800 3 $0.016 $0.048 $0.0022 $0.038 $0.080
🔄 Full Loop
Telegram → Lead lookup → Calendar create → Notion update
Sonnet 18,000 1,500 5 $0.026 $0.077 $0.0036 $0.060 $0.128
📝 Session Closer
Read last 5 sessions → write summary → Notion log
Sonnet 15,000 2,000 3 $0.025 $0.075 $0.0034 $0.058 $0.125
🔍 Research Session
Brave search × 3 + compile + Notion write
Sonnet 35,000 5,000 6 $0.060 $0.180 $0.0083 $0.138 $0.300
📊 Proposal Generator
Transcript + audit + full proposal write
Opus/Sonnet 40,000 8,000 7 $0.080 $0.240 $0.0108 $0.180 $0.400
With prompt caching enabled (cache the 4,500-token base system prompt): costs drop by ~20–35% for short workflows and up to ~50% for the heartbeat. Strongly recommended for production FasClaw.
03

Monthly Cost Estimate — Realistic Usage

Workflow Runs/Month Cost/Run (Sonnet) Monthly (Sonnet) Cost/Run (Haiku) Monthly (Haiku) Cost/Run (GPT-4o mini) Monthly (GPT-4o mini)
⏱ Heartbeat
Every 30min, 24/7
1,440 $0.0072 $10.37 $0.0004 $0.58 $0.0003 $0.43
📋 Telegram Query
~20 queries/day
600 $0.032 $19.20 $0.011 $6.60 $0.0015 $0.90
📌🤖 Calendar Worker
~5 tasks/day
150 $0.048 $7.20 $0.016 $2.40 $0.0022 $0.33
🔄 Full Loop
~2 per day
60 $0.077 $4.62 $0.026 $1.56 $0.0036 $0.22
📝 Session Closer
~1 per day
30 $0.075 $2.25 $0.025 $0.75 $0.0034 $0.10
🔍 Research Session
~2 per week
8 $0.180 $1.44 $0.060 $0.48 $0.0083 $0.07
📊 Proposal Generator
~2 per week
8 $0.240 $1.92 $0.080 $0.64 $0.0108 $0.09
TOTAL / MONTH $47.00 $13.01 $2.14
💡 Optimization insight: The heartbeat is the biggest hidden cost at Sonnet rate ($10/month). Using Haiku for heartbeat + all simple Telegram queries brings monthly total down to ~$12–15. Reserve Sonnet for full loops, research, and proposals. GPT-4o mini is dramatically cheaper for all tasks — the tradeoff is lower quality on complex Thai-language intent parsing.
04

API Cost vs Subscription Plans

Claude Pro
$20
per month
Usage limitStandard
ModelsSonnet, Haiku
Claude Code
Cowork sessions~40–60/mo
API access❌ Separate
FasClaw runs❌ Not covered
Max x5
$100
per month
Usage limit5× Pro
ModelsSonnet, Haiku, Opus
Claude Code
Cowork sessions~200–300/mo
API access❌ Separate
FasClaw runs❌ Not covered
Max x20
$200
per month
Usage limit20× Pro
ModelsAll incl. Opus 4.6
Claude Code✅ Priority
Cowork sessions~800–1000/mo
API access❌ Separate
FasClaw runs❌ Not covered
API Direct (FasClaw)
$13–47
per month (est.)
Usage limitUnlimited
ModelsAny via OpenRouter
Runs 24/7✅ Autonomous
Calendar triggers
Self-hosted VPS
Manual sessions❌ Cowork only
Cost at low usage$2–5/mo (Haiku)
Cost at high usage$40–80/mo
Breakeven: How Many Runs Before API Beats Subscription?
Workflow Cost/Run (Sonnet) Runs to = Pro ($20) Runs to = Max x5 ($100) Runs to = Max x20 ($200) Your est. monthly Verdict
Heartbeat $0.0072 2,778 13,889 27,778 1,440 API wins
Telegram Query $0.032 625 3,125 6,250 600 Even w/ Pro
Calendar Worker $0.048 417 2,083 4,167 150 API wins
Full Loop $0.077 260 1,299 2,597 60 API wins
Session Closer $0.075 267 1,333 2,667 30 API wins
Research Session $0.180 111 556 1,111 8 API wins
Proposal Generator $0.240 83 417 833 8 API wins
⚖️ The Real Verdict
FasClaw API and Cowork subscriptions serve DIFFERENT purposes — they're not substitutes. FasClaw runs 24/7 autonomously triggered by calendar and Telegram; Cowork is your manual deep-work tool. You need both.

Best stack recommendation: Keep Cowork Pro ($20/mo) for manual sessions + FasClaw API with Haiku for heartbeat + simple queries / Sonnet for complex workflows = ~$33/month total vs Max x20 alone at $200/month which still doesn't cover FasClaw API.

Upgrade to Max x5 ($100/mo) only if: you're doing 3+ hours of Cowork sessions daily. Upgrade to Max x20 ($200/mo) only if you're a power builder and want Claude Code priority + Opus 4.6 access without worrying about limits.
05

Runtime Estimates per Workflow

Workflow Haiku Sonnet Opus GPT-4o
⏱ Heartbeat
1–2s 3–5s 6–10s 3–6s
📋 Telegram Query
3–6s 8–15s 15–25s 8–12s
📌🤖 Calendar Worker
10–20s 20–40s 35–60s 20–35s
🔄 Full Loop
30–50s 45–90s 90–150s 40–75s
📝 Session Closer
30–60s 60–120s 2–3 min 50–100s
🔍 Research Session
2–4 min 4–8 min 6–12 min 3–6 min
📊 Proposal Generator
3–5 min 5–10 min 8–15 min 5–8 min

Time breakdown per run includes: model latency + Notion API calls (~200ms each) + Google API calls (~300–500ms each) + Brave Search (~500ms–1s). Bottleneck is usually Notion rate limit (3 req/s max).

Runtime Breakdown (Full Loop, Sonnet)
Model inference
~18s
Notion API calls
~10s
Google Calendar
~5s
Context routing
~4s
Overhead / retry
~3s
VPS latency note: All numbers assume low-latency API connections. Your VPS (hstgr.cloud) adds ~50–100ms round-trip vs US data centers. Acceptable for async workflows, noticeable for real-time Telegram replies.
06

Potential Bugs & Failure Modes

Bug / Failure Severity Affected Workflows Root Cause Mitigation
🔤 Alias Not Resolved
Thai nickname not found in Lead DB Aliases field → workflow stalls or errors
HIGH Calendar Worker
Full Loop
Client name spelled differently in calendar vs Lead DB aliases field Maintain alias list. Add fuzzy match fallback. Return "alias not found" to Telegram with suggestions.
⏱ Notion Rate Limit Hit
Notion API caps at 3 req/sec — multi-step workflows can trigger 429 errors
HIGH All multi-step Sequential tool calls in rapid succession exceed Notion's 3 req/s limit Add 350ms delay between Notion calls. Implement exponential backoff on 429. Batch related queries.
📌 Emoji Parse Fail
📌🤖 not detected if calendar event title is truncated or emoji is different variant
HIGH Calendar Worker Unicode emoji variants (text vs emoji presentation), or long titles truncated by Google Calendar API Use string contains check, not exact match. Test with both U+1F4CC variants. Log all scanned events.
📅 Thai Date Misparse
"พรุ่งนี้", "อาทิตย์หน้า", "วันศุกร์" misinterpreted or wrong timezone
MED Full Loop
Telegram Query
Relative Thai dates require context-aware parsing. VPS timezone may differ from Bangkok time (UTC+7) Always confirm date back to user in Telegram before creating event. Set VPS timezone to Asia/Bangkok.
🔗 Wrong Notion Relation ID
Task created in TASKS DB linked to wrong client ID
MED Calendar Worker
Full Loop
Two clients with similar names — Context Router returns wrong Notion page ID Log resolved client ID + name to Telegram confirmation message. Add "Linked to: [Name]" in reply.
📦 Context Overflow
Large Notion DB response (many records) pushes context over model limit
MED Research
Proposal
Unfiltered DB queries return 100+ records. Large search results from Brave. Always filter Notion queries (max 10 results). Summarize search results before feeding to next step.
💾 Duplicate Entry Created
Calendar cron runs twice → same task created twice in TASKS DB
MED Calendar Worker Cron job overlaps if previous run takes longer than cron interval. No idempotency check. Before creating task, query TASKS DB for existing entry with same calendar event ID. Skip if exists.
🤖 Haiku Misclassification
Haiku heartbeat classifies a complex intent as simple → routes wrong model
LOW All Haiku is less reliable for nuanced Thai intent classification than Sonnet Use Haiku only for yes/no health pings. Route all Telegram messages directly to Sonnet.
🔌 VPS Downtime
Self-hosted VPS goes down → all autonomous workflows stop silently
LOW All autonomous No SLA on self-hosted infrastructure. Network blips, OOM kills, etc. Set up external uptime monitor (UptimeRobot free tier). Alert to Telegram if ping fails for 5+ min.
🔑 API Key Expiry
Anthropic/OpenAI API key expires or runs out of credits silently
LOW All No automatic credit top-up or key rotation alert Set spending alerts in Anthropic console ($10, $30, $50 thresholds). Use auto-recharge.
07

Key Insights & Recommendations

01
Use Haiku for heartbeat only
Switching heartbeat from Sonnet to Haiku saves ~$10/month. That's $120/year for a simple health ping. No quality difference for yes/no checks.
02
Prompt caching = 20–35% savings
The 4,500-token system prompt repeats on every run. Enabling prompt caching brings that from $3.00/1M to $0.30/1M per cached read. Easy win.
03
GPT-4o mini for simple queries
At $0.15/$0.60 per 1M tokens, GPT-4o mini runs Telegram queries 20× cheaper than Sonnet. Worth testing for standard DB lookups that don't need Thai nuance.
04
API always beats subscription for FasClaw
Claude Pro/Max plans only cover claude.ai/Cowork sessions. FasClaw uses API billing regardless of your plan. They're parallel costs, not substitutes.
05
Max x20 is for Cowork heavy users
If you're running 5+ hours of manual Cowork sessions daily, Max x20 at $200/month makes sense. For 1–2 hours/day, Pro at $20/month is sufficient.
06
Alias resolution is the #1 bug risk
Most FasClaw failures happen when a Thai nickname can't be matched to a Notion Lead DB entry. Keep Aliases field updated for every active client.
💰 Recommended Monthly Stack
Cowork Pro ($20/mo) for manual deep-work sessions + FasClaw API with Haiku heartbeat ($0.58/mo) + Sonnet for complex flows (~$12–20/mo) = total ~$33–40/month.

This beats Max x20 ($200/mo) by $160/month and covers both your autonomous agent AND your manual AI work. Upgrade Cowork to Max x5 ($100) only when you're doing 3+ proposal sessions per day in Cowork.
Pricing sources verified March 2026: Anthropic PricingOpenAI API PricingClaude Max Plans