Stop Burning GPT-5.6 Sol. Hire 5 AI Agents + Setup Router and Replace $3000 Agency for $69/Month

---

The problem: you're paying for a Ferrari to drive to the grocery store

Everyone I know upgraded to GPT-5.6 Sol the day it launched. Thirty bucks a month, the top tier, and they aim it at everything: summarizing a newsletter, fixing a typo, writing a two-line reply.

I did the opposite. I downgraded

Sol is $30. Terra is $15. Luna is $6.

Same model underneath, just different limits and speed

So I stopped asking which tier is smartest and started asking how little brain each task actually needs

---

Then I built a router to answer that for me

I drop a job into Telegram, it scores how hard the job is, and it hands it to whichever of my five agents is cheap enough to do it well. Nine times out of ten, that's Terra

My costs went from a $3,000/month agency to $51 in subscriptions, and here is how

---

The system in 30 seconds

Five agents. Not chatbots — agents. Each has a role, a tier, and a job it's not allowed to leave.

Content and Creative both ride the same Terra key - different system prompts but one subscription. That's the trick nobody prices in

$51/month in model access. + $18 VPS. $69 all-in. vs $3,000 agency

TIER     PRICE    THROUGHPUT     BEST FOR
------------------------------------------------
Luna     $6/mo    ~10K req/day   facts, cleanup, search
Terra    $15/mo   ~5K req/day    90% of everything
Sol      $30/mo   ~500 req/day   hard reasoning only
------------------------------------------------
3 keys, shared by 5 agents  =  $51/mo access
+ Hetzner VPS               =  $18/mo
                    ALL-IN  =  $69/mo   vs $3,000 agency

---

The router is the whole thing (and it's not AI)

No LangChain. No CrewAI. Just a function that scores each task 0–100 and picks a tier

def route(task: str) -> str:
    s = 0
    s += 20 if len(task) > 500 else 0
    s += 30 if "reason"    in task else 0
    s += 25 if "strategy"  in task else 0
    s += 15 if "analyze"   in task else 0
    s += 10 if "write"     in task else 0
    s -= 15 if "summarize" in task else 0
    return "luna" if s < 30 else "terra" if s < 85 else "sol"

That's the decision core - 12 lines.

The full script (~150 lines with logging + a Supabase feedback loop that re-scores tiers by failure rate) is in the repo. Runs in ~40ms

---

A day in the life

9:00 AM. Telegram already has my Research digest from Luna: 10 articles → bullets, 3 trends summarized, my leads CSV cleaned. Cost: ~$0.02.

My attention: 3 minutes

10:00 AM. I drop "write a thread on GPT-5.6 routing."

Router scores it 42 → Terra.

Draft back in 40 seconds

2:00 PM. "Should we pivot to B2B given last month's numbers?"

Router scores it 91 → Sol.

Two minutes later: full breakdown with risks

And again - I don't pick the model. I don't write the prompt. I only review what Sol flags

---

Why Terra, not Sol?

Terra isn't a worse Sol

It's a different shape: more throughput, less deep reasoning

My month-one mistake was thinking Sol was better at everything. Then I read my own logs - 847 tasks - and the line was obvious:

• Luna - facts, volume, cleanup. Cheap and fast
• Terra - 90% of real output. The workhorse
• Sol - only the multi-step problems where being wrong is expensive

The stack - what plugs into what

• Telegram bot - my only interface. Drop a task, get a result
• n8n - orchestration, self-hosted, $0
• Python router - FastAPI, on the VPS
• OpenAI API - 3 tier keys (Luna / Terra / Sol)
• Supabase - task history + failure rate per tier (free)
• Hetzner CX31 - $18/mo, runs all of it

No AutoGPT. No agent framework. A router and three keys

---

So Before vs After

                 BEFORE          AFTER
------------------------------------------------
cost / month     $3,000 agency   $69 (AI + VPS)
my ops time      30 hrs/week     4 hrs/week
turnaround       3-7 days        2-4 hours
uptime           office hours    24/7, never quits

---

Build it in one evening

1. Buy Luna + Terra. Skip Sol for week on
2. List your 10 most common tasks.
3. Score each 0–100 by han
4. Route them manually for 7 days. Watch where Terra breaks.
5. Terra breaks → bump to Sol. Sol's overkill → drop to Terr
6. Automate the routing from your logs, not my rubric.

I built mine from 847 real tasks

---

Full setup, free:

• Router code (Python)
• n8n workflow (1-click JSON import)
• System prompts for all 5 agents
• My tier-scoring rubric

Comment "TERRA SWARM" and I'll DM you the repo.

ps. thanks to my friends who inspired me and those who support me :3