What Are Tokens and Why You Pay for Them

You use AI every day. But do you know what a token actually is?

It’s not a word. It’s not a character. It’s the unit that determines how much you pay for every single API request.

Token ≠ Word

LLMs don’t process text like humans. They break everything into pieces called tokens. And each model breaks text differently.

Take the word “development”:

• GPT-4o tokenizes it into 3 tokens: devel + op + ment
• Claude tokenizes it into 2 tokens: develop + ment

Same word. Different token count. Different cost.

This means the same prompt sent to two models at the same price-per-token can cost completely different amounts. The tokenizer is the hidden variable in every AI pricing comparison.

The Real Pricing Table (May 2026)

Here’s what 1 million input tokens costs across major models:

Model	1M Input	1M Output	Best For
Claude Opus 4	$15.00	$75.00	Complex reasoning, code
GPT-4o	$2.50	$10.00	General tasks, balanced
Claude Sonnet 4	$3.00	$15.00	Code, analysis
Gemini 2.5 Pro	$1.25	$10.00	Long context, multimodal
GPT-4o-mini	$0.15	$0.60	Simple tasks, high volume
Claude Haiku 4.5	$0.80	$4.00	Fast responses, triage

Source: AI API Pricing Comparison 2026

The spread is massive. Claude Opus costs 100x more than GPT-4o-mini per input token. Are you using the right model for each task?

Cost Per Token vs Cost Per Result

Here’s where most developers get pricing wrong: they compare price-per-token and pick the cheapest model.

But two models at $5/MTok can cost completely different amounts for the same task. Why?

1. Different tokenizers — same text, different token count
2. Different completion lengths — one model might be more verbose
3. Different accuracy — a cheaper model might need 3 retries to get it right

What actually matters is cost per successful result, not cost per token.

Example: Code Review Task

Model	Tokens Used	Cost	Accuracy	Cost Per Correct Result
Opus 4	3,200	$0.29	95%	$0.31
Sonnet 4	2,800	$0.05	88%	$0.06
GPT-4o-mini	4,100	$0.003	62%	$0.005

For a code review, Sonnet gives you 88% accuracy at 6 cents. Opus gives you 95% at 31 cents. Mini gives you 62% at half a cent — but you’ll spend more time fixing its mistakes.

The right answer depends on what you’re building. Production code review? Use Opus. Linting suggestions? Mini is fine.

How One Endpoint Becomes $2,100/Month

Let’s trace a real scenario:

Single API request:
  prompt:    847 tokens  → $0.012
  response:  2,341 tokens → $0.058
  total:     $0.07 per request

Scale it:
  1,000 requests/day  → $70/day
  30 days              → $2,100/month

And that’s just one endpoint. Most applications have 3-5 AI-powered features, each making hundreds of requests.

The Smart Approach: Route, Cache, Monitor

The developers cutting 60-80% of their AI costs aren’t switching to worse models. They’re doing three things:

1. Smart Routing

Route each query to the cheapest model that can handle it:

// Simple classification → use mini ($0.15/MTok)
// Code generation → use Sonnet ($3/MTok)  
// Complex architecture → use Opus ($15/MTok)

A typical split of 70% mini / 20% mid / 10% premium reduces average cost by 60-80%.

2. Prompt Caching

If your system prompt is 2,000 tokens and you send 1,000 requests/day, you’re paying for those 2,000 tokens 1,000 times. Prompt caching reduces that to nearly zero after the first request.

3. Real-Time Monitoring

You can’t optimize what you can’t see. A proxy that logs every request gives you:

• Cost per model, per project, per day
• Token usage patterns (are you over-prompting?)
• Budget alerts before you overspend

Start With Visibility

Before optimizing anything, answer one question: how much does each of your projects cost per day?

If you can’t answer that, you’re flying blind — just like Uber’s 5,000 engineers who burned $10M+ in four months.

Veltrix Platform gives you that visibility in 5 minutes. One proxy endpoint, all providers, every token tracked.

→ model: claude-opus-4-6
→ tokens: 1,847  
→ cost: $0.032
→ latency: 340ms
→ project: my-saas

Know your costs. Then optimize them.

---

Veltrix Platform — FinOps for AI tokens. Visibility, savings, and governance without changing your code. Get started →