Cost model
Two views — incremental cash cost vs allocated economic cost — plus LLM pricing, illustrative scenarios, the four budget phases and the initial cash cap.
Clawlas tracks cost in two views. The cash view asks how much more it costs this month; the economic view asks whether it would be profitable standing alone. The cost engine, budgets and the routing layer all serve those two numbers.
Cash vs allocated economic cost
The cash view answers “how much more will I pay this month?” The economic view answers “is Clawlas profitable if it had to stand on its own?” Both are tracked; neither is sufficient alone.
- Incremental cash cost — new money spent because Clawlas exists: paid LLM calls, premium market data, RPC overages, X/social usage, extra backups, a future GPU, and trading fees / on-chain costs. The existing Hostinger KVM 4 and shared domain are $0 new cash.
- Allocated economic cost — Clawlas’s share of VPS CPU/memory/disk/network, domain & DNS admin, backup storage, maintenance time and future GPU idle/depreciation. Recorded in shared_infrastructure_allocations even though no new cash is spent, so an honest standalone P&L is always available.
Reference LLM pricing
Per 1M tokens, official public prices at the reference date. These change frequently — treat them as planning estimates and recheck the vendor page before committing.
The open-weight tier — DeepSeek, Qwen and Kimi (Moonshot) — usually wins on quality-per-dollar for high-volume work, and the same weights can later be self-hosted (Phase C), so the router prefers it whenever it clears the per-task quality bar.
| Provider / model | Input | Cached in | Output | Suggested use |
|---|---|---|---|---|
| Gemini 3.1 Flash-Lite | $0.25 | $0.025 | $1.50 | High-volume classification / extraction |
| Gemini 2.5 Flash-Lite | $0.10 | $0.01 | $0.40 | Very cheap simple processing |
| Gemini 2.5 Flash | $0.30 | $0.03 | $2.50 | Standard analysis |
| Gemini 2.5 Pro | $1.25* | $0.125* | $10.00* | Deeper reasoning (*≤200k prompt) |
| GPT-5 mini | $0.25 | $0.025 | $2.00 | Precise structured tasks |
| GPT-5 | $1.25 | $0.125 | $10.00 | Advanced analysis |
| GPT-5.5 | $5.00 | $0.50 | $30.00 | Rare premium analysis |
| Claude Haiku 4.5 | $1.00 | see vendor | $5.00 | Fast analysis & tool use |
| Claude Sonnet 4.6 | $3.00 | see vendor | $15.00 | Deep opportunity / strategy review |
| DeepSeek-V3 | $0.14 | $0.014 | $0.28 | Cheap reasoning & bulk synthesis (open-weight) |
| Qwen3 Flash | $0.19 | see vendor | $1.13 | Low-cost multilingual extraction (open-weight) |
| Kimi K2.5 (Moonshot) | $0.60 | $0.10 | $3.00 | Agentic tool use & long context (open-weight) |
Illustrative monthly cloud-LLM cost
Light = 2.5k in / 300 out on a cheap model; deep = 10k in / 1.5k out on a strong model; report = 30k in / 2.5k out on a strong model. Add ~10–25% for retries, memory compression, evaluation and overhead.
| Scenario | Light/day | Deep/day | Reports/day | ≈ base /mo |
|---|---|---|---|---|
| Prototype | 100 | 5 | 1 | ~$13 |
| Normal personal | 300 | 15 | 2 | ~$37 |
| Intensive | 1,000 | 75 | 4 | ~$158 |
Budget phases
Incremental cash, excluding trading capital, fees, gas and slippage.
Phase A — research & paper trading
| Item | Monthly cash |
|---|---|
| Existing KVM 4 & domain | $0 (allocation tracked separately) |
| LLM APIs | $5–$25 |
| CCXT / Convex self-hosted | $0 (dev & VPS time) |
| DEX Screener / CoinGecko / RPC / X | $0 initially |
| Additional backups | $0–$3 |
| Incremental cash total | $5–$28 / month |
Phase B — limited real trading, one chain
| Item | Monthly cash |
|---|---|
| KVM 4 & domain | $0 |
| LLM APIs (after routing) | $10–$40 |
| Market data | $0–$39 |
| RPC | $0–$50 |
| Swap API / aggregator | $0–$50 |
| X / social | $0–$25 |
| Backups & monitoring | $0–$10 |
| Incremental technical total | $10–$214 / month |
Phases C & D — hybrid models and advanced realtime
| Phase | Driver & range |
|---|---|
| Phase C — hybrid OSS + cloud | Dominated by GPU utilization & premium data; cloud fallback $5–$50 |
| Phase D — advanced realtime multi-chain | $304–$1,299+ / month plus GPU; premium data $199–$499+ |
Phase C is justified only when validated local quality and utilization beat cloud cost per valid result; Phase D only when measured net profit covers premium data, RPC streaming and GPU.
Trade costs dominate, not LLM spend
Net PnL must subtract exchange fees, spread, slippage, price impact, gas, priority fees, failed-transaction gas, withdrawal/bridge fees and MEV — and only then LLM, API and infrastructure cost. Two illustrations:
- CEX: 10 round trips/day × $100 × 2 sides × 30 days = $60,000 monthly volume; at 0.1%/side that is ~$60 in fees alone, before spread and slippage.
- DEX: a $100 buy realizing $94 of tokens is a 6% immediate loss; the exit can be worse. One badly routed trade outweighs a month of model spend.