Cost & Operations

Estimating Your Monthly Claude Bill Before You Commit

You can estimate a Claude workload's monthly cost on the back of an envelope: monthly volume, times tokens per request, times the price per token. The skill is choosing honest assumptions and labeling them as such.

Claude 3P 101 · Updated July 2026 · Unofficial guide

Claude bills on exactly two quantities: input tokens (text you send) and output tokens (text the model writes), each at a per-million-token rate that depends on the model. That means any workload's monthly cost reduces to one formula, applied twice:

monthly cost = requests per month × tokens per request × price per million tokens ÷ 1,000,000, computed once for input and once for output, then summed. Everything in this article is that formula with realistic numbers.

The price list you'll multiply against

List prices per one million tokens, July 2026. Cloud marketplace list prices match Anthropic's first-party prices; committed-use discounts are negotiated with your cloud provider. Estimate against list price and treat any discount as margin for error.

ModelInput (per 1M tokens)Output (per 1M tokens)Notes
Claude Opus 4.8$5.00$25.00Most capable general model; default recommendation
Claude Sonnet 5$3.00 ($2.00 intro through Aug 31, 2026)$15.00 ($10.00 intro)Balanced workhorse
Claude Haiku 4.5$1.00$5.00Fast/cheap; 200K context (others 1M)

The worksheet: five numbers per use case

For each use case, write down five numbers and mark every one of them as an assumption to be tested: (1) requests per month, (2) average input tokens per request, including the system prompt and any documents, (3) average output tokens per request, (4) the model, and (5) a growth or safety multiplier. Volume usually comes from an existing metric (tickets per month, documents processed, emails received). Token counts are the ones people struggle with, so use a rough anchor: a token is around three-quarters of an English word, so a page of text is on the order of several hundred tokens. Better yet, run twenty real samples through the API and read the exact counts from response.usage: an hour of work that replaces your two shakiest guesses with measurements.

Worked example 1: a support drafting assistant on Sonnet 5

Suppose a support team wants Claude to draft replies for agents to review. All figures below are labeled assumptions, not benchmarks. Assume 20,000 tickets per month get a drafted reply. Assume each request sends 1,500 input tokens (system prompt, ticket text, a few knowledge-base snippets) and produces 400 output tokens. Model: Claude Sonnet 5 at standard list rates ($3.00 in / $15.00 out).

Input: 20,000 × 1,500 = 30,000,000 tokens → 30M ÷ 1M × $3.00 = $90. Output: 20,000 × 400 = 8,000,000 tokens → 8M ÷ 1M × $15.00 = $120. Estimated total: about $210 per month at list price, before any caching savings, or $140 at the Sonnet 5 intro rates ($2.00 in / $10.00 out) available through August 31, 2026. Note the shape: output cost exceeds input cost even though four times fewer output tokens were generated, because output tokens cost five times more.

Worked example 2: email triage on Haiku 4.5

Now a high-volume, low-complexity job: classifying inbound email into queues. Assume 100,000 emails per month, 800 input tokens each (instructions plus the email), and just 50 output tokens each (a category label and confidence). Model: Claude Haiku 4.5 ($1.00 in / $5.00 out).

Input: 100,000 × 800 = 80M tokens → $80. Output: 100,000 × 50 = 5M tokens → $25. Estimated total: about $105 per month to triage 100,000 emails, roughly a tenth of a cent per email at these assumptions. The same workload pointed at Opus 4.8 ($5.00 in / $25.00 out) would come to about $525: same arithmetic, five times the price, which is why matching the model tier to the task is the single biggest lever in your estimate.

Rule of thumb: your volume guess will be more wrong than your token guess. Compute a low / expected / high range by varying request volume (say 0.5×, 1×, 3×) and present the range, not a single number. If even the high case is a rounding error in your budget, stop refining and go build the pilot; measured data beats modeled data.

What the envelope leaves out

Four things, all knowable in advance. Conversation history: chat-style apps resend prior turns, so input tokens per request grow with conversation length; budget for the average conversation, not the first message. Prompt caching: if a large prompt section repeats across requests, caching (available on all four platforms) reduces effective input cost, so your list-price estimate is conservative. Retries and evaluation traffic: add a modest allowance for failed-and-retried requests and for your own testing. Hidden growth: the estimate is per use case; the bill is per organization. A tagging and monitoring setup that attributes spend per feature keeps the sum of envelopes honest.

Where to go next

If the two meters are still fuzzy, start with Claude 3P Pricing, Explained Like a Utility Bill. Then see how the two biggest levers work in practice: Cutting Costs with Model Tiering and Prompt Caching: The Single Biggest Cost Lever.