OpenAI and Claude API cost for a real product: three workloads, real monthly bills

How much do the OpenAI and Claude APIs cost for a real product?

For most business workloads the OpenAI or Claude API bill is tens to a few thousand dollars a month. With prices as of October 2026, extracting 20,000 two-page documents costs about $17 a month on a small model and about $330 on a mid-tier one. An operations agent that handles 3,000 tasks with six model calls each costs about $330 a month on a mid-tier model and about $1,700 on a flagship. A writing feature for 10,000 SaaS users costs about $70 to $1,400 a month depending on the tier. Batch processing, caching and routing easy steps to small models cut these bills by 30 to 70 percent.

Price lists per million tokens are easy to find and hard to turn into a budget. This guide prices three workloads we see often in client projects, line by line, so a CTO or founder can map their own feature onto the closest one. If you want to plug in your own numbers, our LLM API cost calculator does the multiplication for nine models.

Which prices does this guide use?

All bills below use the standard list prices published by OpenAI and Anthropic, checked on October 22, 2026, in USD per million tokens. Prices change several times a year, so treat the totals as orders of magnitude and re-check before you commit a budget.

Two rules shape every bill. Output tokens cost five times more than input tokens on every tier, so long answers and long reasoning are expensive. Cached input costs 10 to 40 times less than fresh input, so a prompt that starts with the same instructions every time is much cheaper than one that changes from the first line. For a side-by-side look at the two providers beyond price, see Claude API vs OpenAI API.

Workload 1: document extraction pipeline

A finance or operations team receives invoices, purchase orders and delivery notes. The pipeline reads each document, returns structured JSON with supplier, dates, line items and totals, and a person confirms anything below a confidence threshold.

Assumptions: 20,000 documents a month, two pages each on average. Each request sends about 5,000 input tokens: the page images and text, plus 1,500 tokens of fixed instructions and the output schema, which are cached. Each answer is about 800 tokens of JSON. About 10 percent of documents are re-run after a validation check fails.

That is 70 million fresh input tokens, 30 million cached input tokens and 16 million output tokens a month, plus the 10 percent re-runs.

What this shows:

• The model bill is small next to the work it replaces. Even the mid-tier real-time bill is under two cents per document.
• Most documents do not need a large model. Clean, typed invoices from repeat suppliers extract well on a small model. Send scans, handwriting and unusual layouts to a mid-tier one, and route by a confidence score.
• Extraction rarely needs real time. If results are needed within a few hours, the Batch API halves the bill.

Our guide on AI invoice and document processing covers how such a pipeline is built and validated.

Workload 2: an operations agent that calls tools

An agent handles internal requests: it reads a ticket or email, looks up the customer and order in your systems, decides what to do, updates a record or drafts a reply, and asks a person to approve anything risky.

Assumptions: 3,000 tasks a month, six model calls per task. Each call sends about 9,000 input tokens: 4,000 tokens of instructions and tool definitions that are cached, and 5,000 tokens of conversation and tool results that change. Each call returns about 500 tokens. Add 20 percent for retries, failed tool calls and evaluation runs.

That is 18,000 calls, 90 million fresh input tokens, 72 million cached input tokens and 9 million output tokens a month.

The trap in this workload is reasoning. If the mid-tier model thinks for 2,000 tokens on every call before answering, output grows from 9 million to 45 million tokens and the monthly bill goes from about $333 to about $765. Thinking is billed as output on both providers. Give the agent a thinking budget only on the decision step and keep lookups and formatting without it.

The other lever is tool results. A tool that returns a full customer record of 3,000 tokens when the agent needs four fields inflates every following call in the task. Trimming tool output to what the agent needs often saves more than switching providers.

Small models are cheap here but tend to choose the wrong tool more often on multi-step tasks. Test the small tier on your evaluation set before you rely on it; a common setup uses a mid-tier model for planning and a small one for classification and formatting. Our AI agent development cost guide covers the build side of the same budget.

Workload 3: an AI writing feature inside a SaaS product

A B2B SaaS product adds a feature that drafts or rewrites text for users: a proposal section, a job description, a product listing, a summary of a record.

Assumptions: 10,000 active users, about 13 generations per user per month, so 130,000 requests. Each request sends 3,000 input tokens: 1,800 tokens of instructions and style guide that are cached, and 1,200 tokens of user content. Each answer is about 700 tokens. Add 10 percent for regenerations.

That is 156 million fresh input tokens, 234 million cached input tokens and 91 million output tokens a month.

Here output dominates: on the mid tier, $910 of the $1,245 base bill is output. If your plan sells at $30 per seat, even the flagship is about 2 percent of revenue per user, but the number grows with every feature you add. Three habits keep it under control: cap the output length per feature, let the user choose between a short and a long draft, and keep a monthly cost-per-user line on your SaaS dashboard next to churn and revenue.

What else is on the bill besides tokens?

The token invoice is the visible part. A production AI feature also needs:

For a single business workflow these usually add $100 to $600 a month before engineering time. Engineering time is the line that grows when nobody owns the feature, which is why every AI feature we ship comes with a test set and a cost-per-task metric from day one.

How do you cut OpenAI and Claude API costs?

These measures, in the order we apply them, cut the three bills above by 30 to 70 percent:

1. Route by difficulty. Send easy cases to a small model and escalate on low confidence.
2. Cache the stable prefix. Put instructions, tool definitions and reference data first and keep them identical between calls.
3. Batch what can wait. Nightly reports, bulk extraction and re-scoring go through the Batch API at about half price.
4. Cap output. Set a maximum answer length per feature and ask for JSON instead of prose where a machine reads the result.
5. Budget thinking. Turn reasoning on only for steps that need judgment.
6. Trim context. Summarise long histories, return only needed fields from tools, attach the relevant pages instead of the whole document.
7. Measure cost per task. Log tokens per request with the prompt version, so a change that doubles cost shows up the same day.

Should you choose the provider by price?

Rarely. At the time of writing the small and mid-tier list prices of OpenAI and Anthropic are identical, and the flagships differ mostly in the cached input price. Choose by accuracy on your own evaluation set, by tool-calling reliability for agents and by your data and hosting requirements, then optimise cost inside the chosen provider. Many products we build keep a thin provider layer so a feature can switch models when prices or quality change.

How we estimate AI running costs for clients

We are a software development company founded in 2020 that builds SaaS products, internal systems and AI features for companies in the US, UK and EU. Our own AI resume builder serves 50,000 monthly active users, so cost per user is a number we track on our own product, not only in client estimates.

During scoping we price your feature the way this guide does: one real request counted in tokens, your expected volume, a model tier per step and the costs around the API. An AI feature added to an existing product starts from $10,000 with us, and a fixed quote follows a short scoping call. See our Claude API integration and OpenAI API integration services for how we build it.