LLM pricing looks trivial — a few dollars per million tokens — until request volume turns fractions of a cent into a monthly bill. This calculator prices a workload across eleven current models from OpenAI, Anthropic, Google, and DeepSeek, showing the per-request cost, the total for your volume, and monthly and yearly projections, with every model ranked cheapest-first for the same workload.
How the Cost Is Calculated
Input (prompt) and output (completion) tokens are priced separately, because providers charge very different rates for reading versus generating:
For daily frequency, monthly cost multiplies the daily total by 30.44 (the average month) and yearly by 365; for monthly frequency, yearly is simply 12×. The stored price table (USD per 1M tokens) includes GPT-5.5 at 30, GPT-4o at 10, GPT-4o mini at 0.60, Claude Opus 4.8 at 25, Claude Sonnet 4.6 at 15, Claude Haiku 4.5 at 5, Gemini 3.1 Pro at 12, Gemini 3.5 Flash at 9, Gemini 2.5 Flash-Lite at 0.40, and DeepSeek V3.2 at 0.28.
Worked Example — One Workload, Four Bills
A request with 2,000 input tokens and 500 output tokens, sent 10,000 times:
| Model | Per request | Total | | --- | --- | --- | | GPT-4o | 100.00** | | Claude Opus 4.8 | 225.00** | | GPT-4o mini | 6.00** | | DeepSeek V3.2 | 4.20** |
The arithmetic for GPT-4o: input 2,000 ÷ 1M × 0.005 and output 500 ÷ 1M × 0.005, so a cent a request — which is 10,000 at a million requests. The spread between the cheapest and most expensive row is over 50×, which is why model routing — sending easy traffic to small models and hard traffic to frontier models — is the single biggest lever in production AI budgets. Note how the output share matters: output tokens cost 3–6× input tokens on most rows, so a chatty model with a long system prompt's worth of generation can cost more than a pricier model that answers tersely.
Estimating Your Token Counts Honestly
The estimate is only as good as the token counts. Count real traffic, not hopes: log a week of production requests and average the actual prompt and completion sizes (the AI Token Counter on this site estimates counts from pasted text). Remember the hidden tokens: system prompts ride along on every request, conversation history grows with each turn, tool definitions and retrieved documents inflate input, and reasoning models bill their internal thinking as output tokens. A 2,000-token average input in your test can be 6,000 in production once history and retrieval are included — rerun the calculator with the honest number before committing to a budget.
Limitations
Prices are a static snapshot maintained in the tool — providers reprice often, so confirm against the provider's current pricing page before signing anything. The model excludes prompt-caching discounts (cached input often costs 50–90% less), batch-API discounts, free tiers, fine-tuning and storage charges, retries and failed requests, taxes, and currency conversion. It prices the tokens you enter; if your true token counts differ, the bill differs proportionally.
Frequently Asked Questions
Why are output tokens more expensive than input tokens?
Generation is the compute-heavy direction: each output token requires a full forward pass produced one token at a time, while input tokens are processed in parallel. Providers price that asymmetry in — typically output costs three to six times the input rate.
Which model is cheapest overall?
For the workload above, DeepSeek V3.2 and Gemini 2.5 Flash-Lite anchor the bottom of the ranking — but "cheapest per token" is not "cheapest per solved task." A weaker model that needs retries, longer outputs, or human correction can cost more per successful outcome; use the ranking to shortlist, then benchmark quality on your own traffic.
How do I cut a projected bill that is too high?
In order of usual impact: route simple requests to smaller models, trim system prompts and history, cap output length, turn on prompt caching for repeated context, use batch APIs for non-urgent work, and cache identical or near-identical responses at your own layer so repeat questions never reach the API at all.