AI Token Cost Calculator: Claude vs GPT vs Gemini vs Bedrock

Ask three engineers what a new AI feature will cost to run and you will get three guesses, each built on a different model, a different idea of how long a prompt is, and no idea whether anyone will cache anything. Token pricing is simple arithmetic, input tokens times the input rate plus output tokens times the output rate, and yet it is the cost almost every team estimates wrong, because the arithmetic has to be done per model, the models are priced across a 300-fold range, and the discounts for caching and batch change the answer by half.

The calculator below does the arithmetic for 21 current models from Anthropic, OpenAI, Google, and Amazon Bedrock, with caching and batch applied the way each provider bills them, and shows the result per request, per month, and per year. The page around it explains how token pricing works, where a list-price estimate stops being useful, and how to turn the estimate into a measured number. For the rates themselves and what drives them, see our guide to LLM token costs by model.

How to Use It

• Requests per month: the number of model calls, not user actions. An agent that makes 30 calls per task counts 30.
• Average input tokens: everything you send, system prompt included. A token is roughly three-quarters of an English word.
• Average output tokens: what the model returns. On reasoning models, add an allowance for thinking tokens, which bill as output.
• Cached share: the portion of each request's input that is a repeated prefix, usually the system prompt and fixed context. Cache reads bill at 10 percent of input (2.5 percent on Claude Fable 5.1).
• Batch share: the portion of requests that can wait up to 24 hours for a response. Batch bills at 50 percent across providers.
• Provider filters: hide families you will not use so the spread reflects your real shortlist.

The Rates Behind the Calculator

List prices per million tokens in US regions as of October 2026. The calculator uses these figures; confirm on each provider's pricing page before budgeting, since model lineups changed three times between July and September 2026.

Three things worth noticing. The frontier tier costs exactly twice the flagship tier on both labs, and Claude Sonnet 5 and GPT-6 Sol are priced identically, so the choice between them is quality rather than cost. The same Claude model can cost 50 percent more on Bedrock than first-party, which is the price of AWS billing and governance. And Gemini 3.x Flash is on a promotional rate that doubles at the end of 2026. Our pricing guides for Claude, ChatGPT and the OpenAI API, Gemini, and Amazon Bedrock cover each lineup in depth.

A Worked Example

A support-triage feature handles 500,000 requests a month with about 2,000 input tokens and 300 output tokens each, and 1,500 of those input tokens are a fixed system prompt. With the defaults, no caching and no batch, the calculator shows about $175 a month on GPT-6 Luna, $1,750 on Claude Haiku 4.5, $3,500 on Claude Sonnet 5 or GPT-6 Sol, and $17,500 on Claude Fable 5.1 or GPT-6 Astra, a 100-fold spread for the same traffic.

Set the cached share to 75 percent, since three-quarters of each request's input is the system prompt, and the Sonnet 5 figure falls to about $2,150. Set batch to 50 percent, since half the volume is historical backfill, and it falls to about $1,610. The same feature on the same model now costs less than half the headline estimate, and routing the routine two-thirds of requests to Haiku 4.5 would roughly halve it again. The calculator makes those decisions visible before any code is written; the measurement afterward is a different problem.

What the Calculator Cannot Know

A list-price estimate is a floor, and the gap between the floor and the bill is where AI cost surprises live.

• Calls per action. An agent that plans, searches, reads, retries, and summarizes may make 30 or 40 model calls to complete one user request. The calculator prices calls; the business counts actions.
• Retries and failures. Every retried call bills its tokens again. A brittle tool integration can double the cost of the path it sits on without changing the traffic shape.
• Conversation growth. Each turn in a chat resends the full history as input, so the twentieth message costs twenty times the first.
• Reasoning tokens. Models with extended thinking bill their intermediate reasoning as output. A three-sentence answer can sit on several thousand billed tokens.
• Which model actually ran. Routing, fallbacks, and defaults mean the model in production is often not the one in the estimate.

Those five mechanics are why the State of Tokenomics 2026 survey found three in four enterprises unable to connect AI spend to a business outcome their CFO would accept, even when they could see the total. Estimating is easy; attributing what was actually billed to the feature and team that caused it is the work, as our guide to AI unit economics sets out.

How Opslyft Does This

The calculator tells you what a feature should cost at list price. Opslyft tells you what it did cost, who caused it, and why it changed, by treating AI as another provider in the same FinOps program as the cloud. AI services billed through your clouds, Amazon Bedrock, Azure OpenAI, and Vertex AI, arrive in Explorer with the rest of the AWS, Azure, and GCP bill, and direct provider spend such as OpenAI is ingested alongside them, so token cost appears next to the compute, Kubernetes, and Snowflake spend the same feature runs on, broken down by team, project, service, or customer and tracked over time.

Dimension Studio is where the calculator's estimate becomes a measured unit cost: it maps AI spend to teams, projects, environments, and features by the same rules as cloud spend, splits shared costs such as an embedding pipeline that serves several products, and reports cost per customer and cost per feature, the number that connects a model bill to a margin. Monitor watches token spend continuously and separates normal fluctuation from a real anomaly, attaching a root-cause explanation, the feature, the model, and the change that caused the spike, and routing it to the owner's Slack channel before it compounds; the feedback loop that improves its accuracy is the same one that catches the retry storm the calculator could not predict. Tickets turns the fix, routing a feature to a cheaper model or enabling caching, into an assigned item in Jira, ServiceNow, or Asana with the saving tracked to close. And Iris answers the question finance asks first, which AI features cost the most per customer and why, in plain English. Our guides to FinOps for AI and why AI costs are cloud costs now set out the practice; the platform runs it.

The Bottom Line

Token pricing is arithmetic, and the calculator does it across every model you are likely to consider, with the two discounts that change the answer most. Use it to choose a model and a shape before you build, and to see that routing the routine majority of requests to a smaller model is worth more than any rate negotiation. Then measure. The estimate is the floor; agents, retries, history, and reasoning tokens are the bill, and attributing that bill to the features and teams that generate it is what turns AI from a cost nobody can explain into one that someone owns. That measurement, across AI and cloud together, is what Opslyft's cost visibility is built for.