8 min read Updated Aug 2026 Optimization

How I Cut 90% of My AI Agent Costs

Running three agents 24/7 costs money. Here's how I reduced token burn from ~$50/month to near-zero while increasing output.

The trick isn't cheaper models — it's smarter routing and never repeating context.

Where the money goes

Agent costs break down into three buckets:

Bucket	% of spend	What it is
LLM inference	70%	Chat/completions calls (input + output tokens)
Web search	20%	Exa, Brave, Tavily for research
Embeddings	10%	GrayMatter vector indexing

The 70% bucket is where you win. Smart routing cuts this by 90%.

9router fallback chain

Don't pay for premium models when free ones work. My config:

# Primary: quality model (rate-limited, paid)
primary: nous/tencent/hy3:free

# Fallback chain: free tiers in order of reliability
fallback:
  - custom/openrouter/minimax/minimax-m3:free
  - custom/meituan/longcat-2.0:free
  - ollama/deepseek-v4-flash:0731-cloud  # local, no API key needed

# Synthesis vs routing split
synthesis: openrouter/minimax/minimax-m3:free  # long-form content
routing:   ollama/glm-5:cloud                  # quick decisions

Result: 95% of requests hit free tiers. The 5% that need quality use paid models only when it matters.

Local vs cloud tradeoffs

Local inference (Ollama on VPS) is "free" but has hidden costs:

GPU RAM: 7B model needs 6GB VRAM. My Radeon 780M (512MB) can't run it
Speed: CPU inference is 5-10x slower than API
Quality: 7B models lose to cloud 70B+ on complex reasoning

My rule: Local for STT (whisper), embeddings (bge-m3), and quick classification. Cloud for synthesis, planning, and research.

Free-tier LLM tier list

Ranked by real-world reliability for agent workloads:

Model	Provider	Reliability	Best for
minimax-m3:free	9router/OpenRouter	★★★★★	Synthesis, content, planning
deepseek-v4-flash	9router/Ollama	★★★★☆	Code, reasoning, research
longcat-2.0:free	Nous Portal	★★★★☆	General agent tasks
glm-5:cloud	9router/Ollama	★★★☆☆	Quick decisions, classification
gpt-oss-120b	OpenRouter	★★★☆☆	Complex reasoning (rate-limited)

Ollama quota resets Monday morning. Don't rely on a single free tier.

Monitoring token burn

Track these metrics weekly:

Tokens/agent/day — if one agent spikes, it's looping
Cache hit rate — prompt caching saves 70%+ on repeated prefixes
Fallback frequency — high fallback = primary model rate-limited
Context bloat — trim old context before it costs you

I log every agent run to journal.md with token counts. After 73 sessions, patterns emerge.

Next: How to Run a 3-Agent AI Fleet on One Laptop — the actual hardware and A2A config.