TTokenCost

How far does your budget go?

Most cost calculators tell you what a prompt costs. This one answers the reverse question: given a fixed monthly budget, how many requests can you actually serve — and which model squeezes the most out of it?

With $100 per month at 1K in / 500 out per request:
GPT-5 nano gives you 400K requests
versus 1.9K with Claude Opus 4.1 — a 210x difference.
ModelVendorCost / requestRequests / monthRequests / day
GPT-5 nanoOpenAI$0.00025400K13.3K
GPT-4o miniOpenAI$0.00045222.2K7.4K
DeepSeek V3.2 (chat)DeepSeek$0.00049204.1K6.8K
Llama 4 MaverickMeta$0.00069143.9K4.8K
GPT-5 miniOpenAI$0.0012580K2.7K
Gemini 2.5 FlashGoogle$0.0015564.5K2.1K
DeepSeek R1 (reasoner)DeepSeek$0.0016560.8K2K
Kimi K2Moonshot$0.0018554.1K1.8K
Claude Haiku 4.5Anthropic$0.0035028.6K952
Qwen3 MaxAlibaba$0.0042023.8K793
Mistral Large 2Mistral$0.0050020K666
GPT-5OpenAI$0.0062516K533
Gemini 2.5 ProGoogle$0.0062516K533
GPT-4oOpenAI$0.0075013.3K444
Claude Sonnet 4.5Anthropic$0.0109.5K317
Grok 4xAI$0.0109.5K317
Claude Opus 4.1Anthropic$0.0521.9K63

Does not include prompt caching, batch discounts or retries. Production workloads typically see 10-25% overhead above these figures.

How to use this for real planning

Start with your expected traffic rather than your budget. If you predict 20,000 requests per month at 1,000 input and 500 output tokens, the calculator tells you immediately whether a frontier model is viable or whether you need a mid-tier one. Working backwards from a budget tends to produce optimistic plans.

A useful pattern: route simple queries to a cheap model and escalate only when confidence is low. Most chat products see 70-85% of traffic fall into the cheap bucket, which often cuts total spend by more than half versus a single frontier model.

Leave headroom. Retries, failed generations, system prompts and conversation history routinely add 10-25% to the raw per-request math shown here.