AI Agent Costs Spiraling? A Practical Guide to Cutting From $200/hr to $30/hr
How It Started One day the boss sent a screenshot — daily cost: $1,747! Average of $200+/hour, 99% from claude-opus-4-6. Clearly unsustainable. We spent half…
- Added to the lab
- Downloads
- 0
- Human review
- Human reviewed
How It Started
One day the boss sent a screenshot — daily cost: $1,747! Average of $200+/hour, 99% from claude-opus-4-6. Clearly unsustainable. We spent half a day doing a systematic optimization: dropped from $200/hr to $30-40/hr — an 85% reduction.
Three Main Cost Culprits
Culprit 1: Main Agent Using the Most Expensive Model
The scheduling coordinator was defaulting to claude-opus-4-6, but scheduling work is just read task → decide → send message. Opus is completely overkill. Fix: switch main Agent to Sonnet. Immediate impact.
Culprit 2: Context Too Large (160K Tokens)
System prompts + workspace files + conversation history stacking up, 160K tokens per call. Used /compact to compress:
/compact
# Result: 159K → 21K, down 87%Culprit 3: Heartbeat Model Using Local Ollama
Ollama cold start takes 10 seconds, occupying the main queue and causing disconnections. Switched to Haiku (millisecond response), changed heartbeat from 30 minutes to 4 hours — eliminated pointless wake-ups.
Model Tiering Strategy
Matching model to task complexity is the core of cost optimization:
- Scheduling/orchestration → Sonnet (Opus is unnecessary)
- Content/copywriting → GPT-4o-mini (sufficient for creative tasks, saves 80%)
- PM/QA/ops → Haiku (structured tasks, speed priority)
- Translation/code → Sonnet (quality-sensitive, keep it)
Self-Audit Checklist
- ✅ Is the main scheduler using the most expensive model? Sonnet is enough for scheduling
- ✅ Is context over 60%? If so,
/compactnow — don't wait until it's full - ✅ What model is the heartbeat using? Simple triggers use Haiku, not Sonnet
- ✅ Does the content Agent need a powerful model? GPT-4o-mini is fine for copywriting
- ✅ Are any Agents idle-spinning? Check for high-frequency useless calls
SFD Lab Notes
This is a real pitfall summary. The core is matching model to task complexity: Opus for complex reasoning, Sonnet for most tasks, Haiku for simple triggers, GPT-4o-mini for content creation. Get the match right and costs naturally come down.
Context compression is the most overlooked technique. Carrying 160K history on every call adds up fast even with a low unit price. Make /compact a habit — it's the easiest way to save money.