OpenAI API: My Year on the Platform — What Works and What Doesn't

A year ago, I connected the OpenAI API for a small SaaS project. I thought: “Fifteen minutes and we’re in production.” Reality turned out to be somewhat more complicated.

Website: platform.openai.com

How It Started

The idea was simple: an AI assistant for processing customer inquiries. GPT-4o had just launched, token pricing looked reasonable, and the documentation was easy to read. I registered, got an API key, wrote my first request — and got a response in two seconds. “Magic,” I thought. Then the nuances began.

First Pitfalls: Rate Limits and Billing

The free tier gives you $5 in credits, but once they’re exhausted, the API simply stops responding — it doesn’t slow down, it hits a wall. Rate limits at Tier 1: 500 requests per minute. Sounds like a lot, but if your service makes 3-4 calls per user request (classification → generation → validation), the limit gets eaten up quickly.

Solution: Tier 2 and higher unlock after spending $50 and a week of usage. Limits then jump to 5,000 RPM, and things get easier.

What’s Actually Great

Model variety. This is OpenAI’s main advantage over Claude and Gemini. For every task — a dedicated model:

• GPT-4o — the workhorse. Text, images, audio. The primary work engine
• GPT-4o-mini — cheap and fast. For classification, summarization, simple responses
• o3 / o4-mini — reasoning models. When you need chain-of-thought: math, code, logic
• DALL·E 3 — image generation via API
• Whisper — speech recognition. Still the best on the market for its price
• TTS / TTS HD — speech synthesis with six voices

Function calling. OpenAI implemented external function calling better than competitors. The model independently decides when to call your API, which parameters to pass, and returns structured JSON. Anthropic and Google have equivalents, but they work less reliably.

Vision. The model genuinely understands images — not just describing them, but extracting structured data. We used it for invoice parsing: 96% accuracy on the first run.

Where It Falls Short

Pricing. GPT-4o: $2.50/1M input tokens, $10/1M output tokens. At volumes of 10M+ tokens per month, the bill becomes significant. For comparison, Claude 3.5 Sonnet is about 30% cheaper.

Vendor dependency. You’re tied to one API. If OpenAI changes models (and they do), your prompts may break. Fine-tuning datasets are bound to specific versions. Migrating to another provider isn’t just changing a URL — it’s rewriting prompts.

Localization. Works with Russian, but the quality of complex text generation in Russian is lower than in English. Particularly noticeable on reasoning tasks.

The Numbers: What It Actually Costs

For a SaaS with 1,000 active users and average consumption of 50K tokens per user per month:

• GPT-4o-mini (classification): 50M tokens × $0.15/1M = $7.50
• GPT-4o (generation): 15M input + 5M output = $37.50 + $50 = $87.50
• Embeddings: 10M tokens × $0.02/1M = $0.20
• Total: ~$95/mo for 1,000 users

Reasonable, but the surprise is that as users grow, costs grow non-linearly — because active users consume well above average.

Affiliate Program

OpenAI doesn’t have a direct affiliate program. But there are two paths:

1. Platform partnership. If you’re building a product on top of the OpenAI API and generating significant volume — contact their sales department for partnership terms
2. ChatGPT referral system. OpenAI offers referral links for ChatGPT Plus ($20/mo) through Impact.com. Commission is modest, but conversion is high

Honestly: if your goal is affiliate income, OpenAI isn’t the best choice. But as a tool for building a product that generates income — it’s excellent.

Recommendation

The OpenAI API is a mature platform with the best model variety. For a startup — use it freely, but have a Plan B (backup provider) in case of changes. For enterprise — negotiate custom terms and budget for migration between models.