AI Token Cost Calculator

Tokens are the atomic units LLMs process — roughly 4 English characters per token, but this varies significantly. Common words like 'the' or 'cat' are 1 token. Rare words, code symbols, or non-English text often tokenize into 2–5 tokens per word. A rough guide: 1 page of English prose ≈ 750 tokens, 1 line of code ≈ 10–20 tokens, 1K characters ≈ 200–300 tokens. For exact counts, use tiktoken (OpenAI) or the provider's tokenizer library.

Anthropic's prompt caching lets you mark parts of your prompt (e.g. a large system prompt or reference document) to be cached server-side. Subsequent requests that hit the cache pay only ~10% of the normal input token price for those cached tokens. This is a massive saving if you repeat the same large context — e.g. sending a 50,000-token codebase with every request. Cache hits are charged at $0.30/1M for Claude 3.5 Sonnet instead of $3/1M. OpenAI offers automatic prompt caching for prompts over 1,024 tokens on GPT-4o.

Yes — you pay per token sent and received, not per context window capacity. A 200K context window doesn't cost more than a 128K window if you only send 1K tokens. However, if you do fill a large context (e.g. feeding a 100K-token document), you pay for every input token on every request. Gemini 1.5 Pro has tiered pricing: inputs under 128K tokens cost $1.25/1M, while inputs over 128K cost $2.50/1M.

Yes. OpenAI's Batch API offers 50% off for asynchronous requests with up to 24-hour turnaround — ideal for offline processing. Anthropic offers a similar Message Batches API at 50% off. Google has tiered pricing where higher usage can unlock lower rates. For very high volumes, you can also negotiate enterprise contracts directly with providers. This calculator shows standard on-demand pricing; batch pricing would halve the estimates.

Several costs are often overlooked: (1) Retry costs — failed requests still charge tokens if the model started generating. (2) System prompt overhead — every request includes your system prompt tokens. (3) Tool call tokens — function definitions and call/response cycles add significant token overhead. (4) Streaming doesn't reduce cost. (5) Rate limit retries waste latency. (6) Egress/compute costs if you're proxying through your own infrastructure. Budget an extra 10–20% for these hidden overheads.

Use cheaper models (GPT-4o mini, Gemini Flash, Claude 3.5 Haiku, Llama 3.1 8B via Groq) for: classification, simple extraction, summarization, chatbot responses, translation, and high-volume pipelines where accuracy isn't critical. Use frontier models for: complex reasoning, code generation, long-context analysis, and tasks where errors are expensive. A hybrid approach — route simple queries to cheap models and complex ones to frontier models — often cuts costs 60–80% with minimal quality loss.