Phase 1 Retrospective: AI Models, APIs, and the Cost of Figuring It Out
Date: 2026-03-13
Status: Closing Phase 1 — AI-Assisted Local Infrastructure + Grid Trading
Next: 1-year project with m900
What We Closed
Two projects that ran in parallel since early February 2026 are now transitioning to steady state:
- AI-Assisted Local Infrastructure — setting up m900 (Lenovo ThinkCentre M900 Tiny) as a persistent AI agent running OpenClaw, with memory, crons, and tool access to the real world.
- Algorithmic Grid Trading (EVM + Solana) — deploying and tuning 4 grid bots (Arbitrum, Base, Linea, Solana + a Hyperliquid perp short) with ATR-based dynamic spacing.
Both worked. Neither worked cleanly.
AI Model Comparison: What Actually Happened
Tested across multiple models over ~6 weeks:
| Model | Verdict |
|---|---|
| Claude Sonnet 3.5 / 4 | Best reasoning + tool use. Default choice. |
| Claude Opus | Noticeably better on complex tasks. Expensive. |
| Google Gemini | Good. Already have the GCP environment — should use more. |
| Venice AI | Interesting privacy positioning, but contract quality is poor. Not reliable enough for production. |
| Others | Tested, discarded. |
The real constraint isn't quality — it's queries per user license. Claude Sonnet and Opus are the best tools, but a standard license throttles you fast if the agent is doing proactive work (heartbeats, crons, monitoring). The fix isn't switching models — it's diversifying so Claude handles reasoning-heavy tasks and Google handles volume or batch work.
APIs cost money. Spent ~€300 total across AI APIs (Google being the largest chunk). The lesson: APIs are fine for experimentation, not for ongoing agent overhead. The agent's background tasks should use zero-cost paths wherever possible — system cron + bash > OpenClaw cron + LLM for anything that doesn't require reasoning.
Grid Bots: What the Data Says
4 bots live across 3 EVM chains + Solana. Summary at closing Phase 1:
- Arbitrum Grid — operational, ETH volatile in February, grid held
- Base Grid — smaller capital, stable
- Linea Grid — largest USDC position, profitable in sideways markets
- Solana Grid — added late, 18 trades/day average, currently +4.1%
- Hyperliquid Short — ETH perp 2x, partially hedges EVM grid exposure
ATR-based dynamic spacing (implemented W10) improved the bots' behavior in volatile periods — wider grids on high volatility, tighter on low. Running hourly.
The decision going forward: bots run on system cron, fully automated. The AI agent does not manage bot operations — that stays human-controlled to avoid overloading the agent with tasks that don't require reasoning.
What Didn't Work
- Google Coral TPU on M900 — hardware keying mismatch. The TPU requires M.2 NVMe slot, not WiFi slot. Physical blocker, not software.
- VPS agent (Hetzner) — ran a second agent for basketball scheduling. Deprecated. Unnecessary complexity.
-
Over-relying on AI for simple automation — anything that can be a bash script should be a bash script. The agent's value is in judgment, not in running
df -hon a schedule.
The 1-Year Project: m900 + AI × Blockchain
Starting now. No rush, no sprints. One machine, one agent, one year.
What m900 does:
- Maintains bot rhythm (automated, human-supervised)
- Monitors infrastructure (kernel, security, disk)
- Participates in Moltbook — reads, comments when relevant
- Runs structured tests at the intersection of AI and blockchain interactions
What m900 doesn't do:
- Manage bots directly (human controls bot config changes)
- Burn tokens on tasks that don't require reasoning
- Overextend into tools and integrations that add noise
The working hypothesis: a persistent AI agent on a €150 mini PC, running for 12 months with minimal API spend, produces more durable value than 6 months of heavy API usage and feature churn.
We'll see.
Cost Tally (Phase 1)
| Category | Approx. Cost |
|---|---|
| AI APIs (Google, Claude, Venice AI, others) | ~€300 |
| Crypto bot gas (EVM chains) | ~$15 |
| Hardware (M900 Tiny) | Already owned |
| VPS (Hetzner) | ~€10/month × 4 weeks |
| Total | ~€350 |
The €300 in APIs bought us: a working mental model of which models are worth using, which API surfaces are reliable, and where automation can replace inference.
That's a reasonable education cost. Not repeating it.
What's Next
- Steady-state bots (human-managed configs, AI-monitored)
- Monthly build-log entries (not weekly — weekly was too much overhead)
- Focused AI × Blockchain tests, documented when there's something worth saying
- No new integrations unless they solve a real problem
Top comments (2)
The "APIs cost money" section really resonates. That €300 lesson is one most of us learn the hard way. Your point about diversifying models based on task complexity is smart though.
One thing that helped me was getting instant visibility into what each provider is actually costing. Running Claude for reasoning tasks and Gemini for batch work is a solid strategy, but only if you can see the split in real time. I keep a menu bar token counter running that breaks down cost per provider (OpenAI, Claude, Gemini, OpenRouter etc) and it makes those budgeting decisions way more concrete than checking dashboards after the fact.
Some comments may only be visible to logged-in visitors. Sign in to view all comments.