Your Classify Task Doesn't Need a 500-Token Essay

You asked GPT-4o to classify a support ticket. The correct answer is one word: billing.

Instead, you got 347 tokens explaining why it's a billing issue, what billing means in context, and three follow-up suggestions you didn't ask for.

At $10 per million output tokens, those 346 extra tokens cost 28x more than the one you needed. Multiply by 10,000 calls a day, and you're burning $100/month on AI opinions nobody reads.

The Pattern Nobody Talks About

We analyzed waste patterns across AISpendGuard workspaces and found the same problem everywhere: concise tasks produce verbose outputs.

The last row is the worst offender. Embedding tasks should produce zero text output — the value is in the vector, not the response. Yet many implementations generate text alongside the embedding, paying for tokens that go straight to /dev/null.

Why This Happens

LLMs are trained to be helpful. When you ask "classify this ticket," the model wants to explain its reasoning. Without explicit constraints, it will:

1. State the classification
2. Explain why it chose that label
3. Offer confidence scores you didn't ask for
4. Suggest related categories
5. Add a disclaimer about edge cases

Each of those steps costs output tokens — the most expensive tokens in every provider's pricing.

What AISpendGuard Now Detects

We shipped Rule 9: Output Verbosity — a waste detection rule that flags when concise task types produce disproportionately verbose output.

Here's what it checks:

• classify and route tasks averaging more than 50 output tokens per call
• eval tasks averaging more than 100 output tokens
• extract tasks averaging more than 200 output tokens
• embed tasks producing any text output at all

When the rule fires, you get:

• Severity — how far above the threshold your outputs are
• Estimated savings — the dollar amount you'd save per month by constraining output
• Actionable fix — specific recommendations for your model and task type
• Deep-link filters — click straight to the affected events in your dashboard

No other cost monitoring tool does this. Helicone and Langfuse show you token counts. We tell you which counts are wrong — and what to do about it.

How to Fix It (3 Approaches)

1. Set max_tokens explicitly

The simplest fix. If your classify task needs one word, tell the model:

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": f"Classify this ticket: {ticket}"}],
    max_tokens=10  # One word + safety margin
)

2. Use structured outputs

Force the model to return JSON with exactly the fields you need:

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": f"Classify this ticket: {ticket}"}],
    response_format={"type": "json_object"},
    max_tokens=20
)
# Returns: {"category": "billing"} — not a 347-token essay

3. Add response format constraints in the prompt

Respond with ONLY the category label. No explanation. No reasoning.
Valid labels: billing, technical, account, feature_request, other

The Numbers

A typical classify workload doing 10,000 calls/day on GPT-4o:

That's not a rounding error. That's your margin.

Start Detecting Output Waste — Free

AISpendGuard's free tier (50,000 events/month) includes all 9 waste detection rules, including output verbosity. Set up the SDK in under 5 minutes, send your events, and we'll tell you exactly where your outputs are too verbose — and how much you'll save by fixing them.

No prompts stored. No model outputs recorded. Tags only.

Start tracking for free →

AISpendGuard is the simplest way for dev teams to find and fix wasted AI API spend. Privacy-first, EUR pricing, EU-hosted. Free tier, no credit card required.