The Formula
Monthly cost = requests per day × 30.44 × (avg input tokens × input price + avg output tokens × output price) ÷ 1,000,000. Prices are per million tokens, which is how every major provider quotes them.
Example: 1,000 requests/day at 2,000 input + 500 output tokens on a $2/$10 model costs (2,000×2 + 500×10) ÷ 1M = $0.009 per request, or about $274/month. The same workload on a $0.30/$1.20 budget model runs about $40/month — which is why the comparison table exists.
Why Output Tokens Dominate the Bill
Output tokens are consistently priced 3–6× higher than input tokens across every provider, because generating text is far more compute-intensive than reading it. The practical consequence: two workloads with identical token totals can cost wildly different amounts depending on where the tokens sit.
A summarization job (long input, short output) is cheap. A code-generation agent (modest input, thousands of output tokens per turn, re-sent as input every turn) is expensive. Before optimizing anything else, find out which side of the ledger your workload lives on — changing the model is usually worth less than capping output length or trimming context.
Four Levers That Cut the Bill 10–100×
- Prompt caching. Providers discount cached input tokens by 75–90%. Agent workloads that resend the same system prompt and context every turn are the biggest winners — the cache line in provider price tables applies to exactly this pattern.
- Batch APIs. Non-urgent jobs (overnight summarization, bulk classification) get ~50% off at most providers in exchange for hours of latency. If it doesn't need to be interactive, it shouldn't be interactive.
- Model routing. Route the easy 80% of requests to a budget model and escalate the hard 20%. Most production stacks do this — it's usually the single largest saving after caching.
- Context trimming. Input tokens you re-send "just in case" are billed every single request. Trimming a 10K-token context to 2K cuts the input side of the bill 5× outright.
For live price checks, the calculators above are editable — pull today's numbers from the OpenAI, Anthropic, and Google pricing pages, or an aggregator like aitokenprice.com.
Common Questions
How is the monthly cost calculated?
Monthly cost = requests per day × 30.44 × (average input tokens × input price + average output tokens × output price) ÷ 1,000,000. Prices are per million tokens, the unit every major provider quotes. Edit any preset price to match a provider's official pricing page — they change frequently.
How many words is a token?
For English text, a token is roughly 4 characters or 0.75 words — 1,000 tokens is about 750 words. Code tokenizes less efficiently (symbols and indentation split into more tokens), and non-Latin scripts vary by tokenizer. For estimation, treat a typical support answer or email draft as 300–800 output tokens.
Do the prices include caching or batch discounts?
No — presets are standard-tier, uncached prices. Prompt caching discounts cached input tokens by 75–90% at major providers, and batch APIs take about 50% off in exchange for hours of latency. Workloads that resend the same context every turn (agents, chat with long system prompts) can see their real bill come in far below this estimate once caching is on.
Related Calculators
- Compound Interest Calculator — What saving $200/month on API spend compounds to
- ROI Calculator — Evaluate the return on an AI feature build
- Break-Even Calculator — How many subscriptions cover the API bill