Monthly AI bill simulator
Estimate the monthly bill of running a feature on GPT-5.5, GPT-5 mini, Claude Opus 4.8, Claude Sonnet 4.6, Claude Haiku 4, Gemini 3 Pro and Gemini 3 Flash. Enter the average input tokens per call, the average output tokens, the volume per month and an optional prompt-caching ratio. The page returns the cost per call and per month for each model, ranks them, projects a yearly total and shows when a smaller model gives the same quality at one fifth the price.
Prices reflect 2026 published list rates. Batch APIs and enterprise contracts can reduce these figures significantly. Cached input is billed at roughly 10% of the standard input rate by most providers.
What an AI cost calculator does before you ship a feature
The unit cost of an AI feature is rarely intuitive. A single call to GPT-5.5 or Claude Opus 4.8 may cost a fraction of a cent, which feels free; multiplied by a million calls a month, the same feature lands on a finance review at five-figures monthly. The variables that drive the bill are simple but compound quickly: input tokens per call (which include the system prompt, the retrieved context, the conversation history and the user message), output tokens per call (which often cost three to five times more than input), the call volume per month, and the percentage of input that benefits from prompt-caching discounts. This calculator turns those four inputs into a clean side-by-side comparison of the seven main 2026 models, so you can decide before you ship whether the feature should run on a flagship like GPT-5.5 or Opus 4.8, on a mid-tier like Sonnet 4.6 or Gemini 3 Pro, or on a small model like Haiku 4, GPT-5 mini or Gemini 3 Flash.
The page is also a sanity check for cost engineering. Many AI features start on the most capable model during prototyping then never get retested on smaller models. By running your real input and output sizes through the table here, you can see whether moving from Opus 4.8 to Sonnet 4.6 cuts the bill by five, whether enabling prompt caching at 80% halves the input cost, or whether the workload is so heavy on output tokens that picking a model with a lower output price matters more than picking a cheap input price.
How AI billing actually works in 2026
Generative AI APIs in 2026 follow the same pattern OpenAI introduced in 2020: pay per token, with a different rate for input and output. Input tokens are everything you send to the model: the system prompt, the chat history, the retrieved knowledge chunks for retrieval-augmented generation, the function definitions, the user message and any examples. Output tokens are everything the model returns, including reasoning traces if you ask for them and the final answer. Output is consistently more expensive because generating a token takes more compute than reading one. Most vendors now offer three additional levers: cached input (recently-seen tokens billed at a fraction of the standard rate), batch APIs (asynchronous jobs at a 50% discount), and reserved capacity for high-volume customers.
- Count the tokens you will send: system prompt, history, retrieved context and user message. The total is what you put in the “input tokens per call” field.
- Estimate the tokens the model will return: a short classification answer is 5-30 tokens, a chatbot reply is 100-400, a structured JSON output is 200-2000, an article rewrite is 500-3000.
- Multiply by the call volume: monthly active users, automation runs, scheduled jobs, retries; everything that triggers a call counts.
- Apply the cached input share: if 80% of the input is a stable system prompt and a stable RAG context, prompt caching can drop that share to roughly 10% of the normal input price.
- Compare across models: the same workload may cost $30 a month on Haiku 4, $180 on Sonnet 4.6 and $300 on Opus 4.8 standard (or $600 on Opus 4.8 fast mode for 2.5x throughput). The right model depends on whether the quality or latency difference justifies the gap.
Common use cases for the calculator
- Budgeting a new AI feature. Before signing off on a roadmap item, multiply expected calls by expected tokens for each candidate model. The finance review goes much faster with a single page that shows monthly and annual numbers for every viable model.
- Choosing between flagship and mini. The price gap between Opus 4.8 standard and Haiku 4 is roughly 10x (Opus 4.8 trimmed the historical 30x premium when it launched on May 28, 2026). If the task is short classification, routing or simple drafting, the smaller model is still usually the right pick. The calculator makes the gap concrete instead of abstract.
- Sizing prompt-caching impact. Cached input pricing is a major lever in 2026. Enter your cache ratio (often 70-90% for stable RAG systems) and see how much it cuts the bill for each model. Vendors with the steepest cache discounts (Anthropic, OpenAI) become noticeably cheaper at high cache ratios.
- Comparing reasoning models versus standard. Reasoning-heavy modes (long chain-of-thought, agent loops) use far more output tokens than a normal chat. Run the same workload at 200 output vs 2000 output and watch the bill flip; some workloads are clearly viable on Sonnet 4.6 with reasoning but ruinous on Opus 4.8.
- Planning a year-end migration. If a feature currently runs on Opus 4.8 standard but Sonnet 4.6 reaches the same quality bar, the annual savings table tells you how much budget the migration frees up. With Opus 4.8 down to roughly 1.7x Sonnet (versus 5x on older Opus releases), the migration only pays back if the workload is volume-heavy or the quality gap is genuinely irrelevant.
- Pricing your own product. When building a SaaS that wraps an AI API, the calculator gives you the per-call cost. Pricing the feature at 3-5x the model cost is a common starting point that the calculator makes easy to verify.
Limitations and accuracy notes
This calculator returns a list-price estimate, not an invoice. Actual bills can be lower because of negotiated rates, committed-use discounts, batch APIs, fine-tuned model rebates and free monthly tiers, or higher because of retries, function-calling overhead, image and audio token surcharges, and longer-than-expected outputs. The tokens per call you enter are an average; real workloads have a long-tail distribution, and a 10% spike on output tokens can move the bill by 5-15%. For production budgeting, run the calculator on three scenarios (baseline, plus 20%, plus 50%) and use the high case for capacity planning. The 2026 prices baked into this page are the ones publicly announced as of the publication date; if a vendor changes its pricing, the comparison will need a refresh.
The page runs entirely in your browser. No information about your workload is sent to PeopleAreGeek or any third party. You can paste real volumes, prototype costs and confidential planning numbers without a network round trip. The cost computation is a few multiplications: input tokens times input price minus the cache discount, plus output tokens times output price, times call volume.
Frequently asked questions
Why is the output price always higher than the input price?
Generating a token takes more compute than reading one. The model must run the full forward pass to produce each output token, which means the GPU time per output token is higher. All major vendors price output between three and five times the input rate, so a chat feature dominated by long replies costs more than a search feature dominated by short ones.
What is prompt caching and how do I model it?
Prompt caching lets you pay roughly 10% of the standard input price for tokens the vendor has recently seen. Useful when a large system prompt or RAG context is reused across calls. To model it, estimate the share of your input that is stable (often 60-90% for production RAG systems) and enter it in the “Cached input %” field.
Should I always pick the cheapest model?
No. Cheaper models have lower quality on hard tasks. A classifier or a router can run on Haiku 4 or Gemini 3 Flash. A coding agent, a structured extraction or a customer-facing chatbot usually needs Sonnet 4.6, GPT-5.5 or Opus 4.8. Run quality evaluations on real samples and pick the cheapest model that meets the target accuracy.
How accurate are the published 2026 prices baked into this calculator?
They reflect the publicly announced list price of each vendor at the publication date. Enterprise customers often have negotiated rates. Batch APIs cut prices by half. Cached input cuts the relevant share by 90%. The calculator shows the unadjusted list price; apply your contract or batch discount on top.
Why does the same workload sometimes look cheaper on Gemini and sometimes on Claude?
It depends on the ratio between input and output. Gemini 3 has a lower input price than Claude in 2026; Claude has competitive output. Workloads heavy on input (large RAG context, short answer) favour Gemini. Workloads heavy on output (long generation, agent loops) sometimes favour Claude or GPT-5.5. The ranked tab shows the winner for your specific ratio.
Is the calculation data sent anywhere?
No. All multiplications happen in your browser. The volumes you enter, the cache ratio and the chosen presets stay on your machine. You can use the calculator for sensitive financial planning without exposing the numbers.
Related tools and resources
Cost is one dimension of AI feature engineering. The tools below help size context, plan prompts, and decide what to send to the model in the first place.
