Compare AI API Costs

A token is the basic unit of text that language models process. One token is roughly 4 characters of English text, or about 0.75 words. The word "calculator" is 2 tokens; "AI" is 1 token. Most providers count tokens using the same BPE (Byte Pair Encoding) tokenizer approach, though the exact tokenization varies by model family. When you send a request, every token in your prompt (input) and every token the model generates (output) is counted and billed.

Generating output tokens requires the model to perform inference one token at a time, which is computationally intensive and sequential — each token depends on the previous one. Reading input tokens can be processed in parallel. As a result, providers typically charge 3–10x more for output tokens than input tokens. This means workloads like content generation (long outputs) cost proportionally more than data extraction tasks (short outputs).

Prompt caching lets you reuse expensive input context — like long system prompts, documents, or conversation history — across multiple requests without re-billing the full input cost each time. Once a prompt prefix is cached, subsequent requests using that same prefix pay a fraction of the standard price (typically 10–50% of the normal input rate). For use cases with large, repeated context like RAG systems or document analysis, prompt caching can reduce input costs by 60–90%.

The Batch API processes requests asynchronously and returns results within 24 hours, rather than responding in real time. In exchange for accepting this delay, providers like OpenAI, Anthropic, and Google offer a 50% discount on both input and output costs. Use the Batch API for offline workloads: bulk document processing, nightly report generation, large-scale classification pipelines, or any task where users don't need an immediate response.

Your monthly cost = (average input tokens per request × input price per million ÷ 1,000,000 + average output tokens per request × output price per million ÷ 1,000,000) × monthly request volume. For example, at 1,000 input tokens and 500 output tokens per request, 100,000 requests/month, with a model at $3/M input and $15/M output: ($3 × 1,000/1,000,000 + $15 × 500/1,000,000) × 100,000 = ($0.003 + $0.0075) × 100,000 = $1,050/month. Our calculator handles this math automatically for any combination of models.

It depends on your use case. The cheapest models by raw price per token change frequently as providers compete — use our calculator to compare real costs for your specific input/output ratio, volume, and quality requirements. Budget-tier models typically start under $0.50/M input tokens, while premium flagship models range from $2–15/M input. Mid-tier models offer the best price-performance ratio for most production workloads.

Yes — you are charged for every token in the context window that the model processes on each request, including previous conversation turns, retrieved document chunks, and system prompts. This means multi-turn conversations grow more expensive over time as the history accumulates. To manage this, you can truncate conversation history, summarize older turns, or use prompt caching to reduce the per-request cost of repeated context.

AI API pricing changes frequently — sometimes multiple times per year. Competition between providers has driven prices down dramatically. GPT-4-class capability that cost $60/M output tokens in 2023 now costs under $5/M with current-gen models. We verify all pricing data weekly against official provider documentation. We recommend re-running your cost comparison quarterly to catch price drops and new model releases.