Start with juggernauts

If you ship your Q1 AI feature on Opus or GPT-5, you have done the right thing — not the wrong one. This page exists because the "open-source LLMs and SLMs" pitch reads, to anyone who has ever shipped under deadline, like an implicit "and you should not be using Anthropic". We are not saying that. Anthropic and OpenAI shipping is fine. We adore them. The point is to make sure you always have the option to leave them when their pricing or priorities change.

What's the situation?

You have one quarter. Your boss expects an AI feature in the product by quarter-end. Your performance review depends on it. You have a budget measured in hundreds of dollars per month, not thousands. You have no research team. You will not get one.

Reach for the cheapest open model first and you will burn two weeks debugging context windows, one week tuning prompts, and ship a triage agent that flags ham as spam in 8% of cases. By Q1 close you have a beta nobody trusts and a slipping timeline.

Reach for Opus or GPT-5 first and you ship a working triage agent in 10 days. Quality is a known quantity. The bill is real, but the bill is also the reason your boss says yes to Q2 work — "the feature works; can we make it cheaper?" is a much easier conversation than "the feature does not work yet; can we have more time?"

What's the recommended path?

The Sagewai Training Loop is structured around exactly this arc:

Loading diagram...

Each quarter is a checkpoint, not a phase change. The work you do in Q1 is what makes Q3 possible — the Curator captures every juggernaut response as training data, automatically, without changing your application code.

Show me a runnable thing

Here is the smallest possible juggernaut-bootstrap loop:

import asyncio
from sagewai.core import Agent

async def main():
    agent = Agent(
        name="support-triage",
        model="claude-opus-4-7",  # The juggernaut
        system_prompt="Categorise the support ticket as: refund, "
                      "technical, account, or other.",
    )
    result = await agent.run("My order #1234 never arrived")
    print(result.output)

asyncio.run(main())

Three lines that matter, plus an await. To swap to GPT-5 later, change one string:

agent = Agent(name="support-triage", model="gpt-5", ...)

To run the same code on local Ollama (Q3, after the fine-tune), one more swap:

agent = Agent(name="support-triage", model="ollama/my-finetuned-llama:latest", ...)

That is the LLM-agnostic surface. Application code does not change across the arc. See Example 18 — local LLM routing for the full swap demonstration across Claude, GPT, and Ollama.

Capture as you go

The Q3 cost-down does not work if you have not been capturing training data the whole time. Sagewai's Curator does this automatically: every successful agent run becomes a candidate training sample, scored by user feedback (thumbs-up / thumbs-down) or by an LLM-judge classifier you configure.

from sagewai.curator import Curator

curator = Curator(
    project_id="acme-support",
    quality_filter="user_rating >= 4 AND human_override == False",
)

By Q3, the Curator has captured several thousand high-quality input-output pairs from your Opus runs — more than enough to fine-tune a 3B local model that hits 90%+ of Opus's quality on your specific task at 1% of the cost-per-call.

For the full Curator → JSONL → Unsloth → Ollama loop, see Example 36 — autopilot training loop and Example 38 — Unsloth fine-tune.

What would I do next?

  1. Ship Q1 on Opus or GPT-5. Stop second-guessing the choice.
  2. Wire the Curator from day one — it costs nothing and it is the reason Q3 will work.
  3. Wire the Observatory cost dashboard — by Q2 you will need to point at it during the CFO conversation.
  4. Read Free CUDA via Colab ahead of Q3 so you already know how the fine-tune works before the cost-down ask lands.

Anti-patterns

  1. Optimising before you have a product. The cheapest LLM is the one that ships your feature on time. That is almost never the smallest open model in week one.

  2. Skipping the Curator. Training data has to exist before you can fine-tune. Capturing retroactively from logs is a week of pain; capturing prospectively from the Curator is a one-line wire-up.

  3. Treating Q3 cost-down as a 10-week project. With the Sagewai Training Loop, the same engineer can ship an Unsloth fine-tune on a free Colab T4 over a weekend. See Free CUDA via Colab.

Cross-references