OpenAI API Pricing Explained (2026): Cost per Token & Real Examples

OpenAI Pricing at a Glance

Understanding OpenAI's pricing model is essential for anyone building with the API. Costs are driven by tokens — both input (what you send) and output (what the model generates) — with significant savings available through caching and the Batch API.

Batch API: Save 50% on inputs and outputs for jobs that complete within 24 hours. Ideal for summarisation, classification, and other non-real-time workloads.

How Token Billing Works

OpenAI charges separately for three token types:

• Input tokens: Your prompt, system message, and any context you send. Billed at the standard input rate.
• Cached input tokens: When you use prompt caching, repeated context (e.g. long documents, RAG chunks) is billed at a much lower rate once cached. Cache hits are cheaper than fresh input.
• Output tokens: The model's response. Output is typically 5–10x more expensive than input, so verbose outputs drive cost quickly.

The ratio of input to output matters. A chatbot with 1,000 input tokens and 250 output tokens per turn has a 4:1 input-to-output ratio. Agentic workflows with long tool outputs can flip that ratio and become output-heavy — and expensive.

Cost Calculator Formula

Use this formula to estimate monthly cost:

monthly_cost = (input_tokens / 1_000_000 * input_rate) + (output_tokens / 1_000_000 * output_rate)

// Example: 100M input + 25M output on GPT-5.2
// (100 * 1.75) + (25 * 14) = 175 + 350 = $525/month

For cached input, substitute the cached rate for the portion of input that hits the cache. Batch API cuts both input and output rates by 50%.

Three Real Workloads

1. Chatbot (1,000 in / 250 out per turn)

High volume, moderate context. Caching helps if you reuse system prompts or RAG chunks. GPT-5 mini is often sufficient for simple Q&A; GPT-5.2 for complex reasoning.

2. Summariser (5,000 in / 500 out per doc)

Input-heavy. Batch API is ideal — summaries don't need real-time latency. Caching document chunks across batches can further reduce cost.

3. Agentic Workflow (2,000 in / 1,500 out per step, 5 steps)

Output-heavy. Tool results and long reasoning add up. Consider smaller models for routing and tool use; reserve flagship models for final synthesis.

Cost Levers

• Caching: Cache system prompts, RAG chunks, and repeated context. Cached input is ~10x cheaper than fresh input.
• Batch API: 50% off for non-real-time jobs. Run summarisation, classification, and extraction in batch.
• Smaller model: GPT-5 mini is 7x cheaper than GPT-5.2 on input and output. Use for simple tasks.
• Shorter outputs: Set max_tokens and use structured outputs to avoid runaway generation.

Common Mistakes

• Context explosion: Sending full conversation history without truncation. Summarise or drop old turns.
• Verbose system prompts: Long system messages are billed as input every request. Keep them lean or cache them.
• No truncation: Failing to limit output length leads to unnecessary tokens. Always set reasonable max_tokens.

Monthly Workload Example

Assume 100,000 chats/month, with 1,000 input tokens and 250 output tokens per chat on average:

• 100M input tokens + 25M output tokens monthly

Async batch version: ~50% less if your workload is eligible for the Batch API.