If you’re estimating the Claude API cost for a production feature, the per-model prices below give you the raw numbers—but the two structural forces that will shape your actual bill are the output-token multiplier and the expiration of introductory pricing on Sonnet 5.

(Prices verified July 2026 against Anthropic’s official page; the LLM API pricing calculator that powers these figures stamps every report with PRICING_AS_OF so the table can never quietly go stale.)

Claude API Cost: Per-Model Pricing Breakdown

Anthropic’s standard (non-batch) list prices as of July 2026. Every row was validated against platform.claude.com/docs/en/about-claude/pricing on that date.

Two things jump out immediately: output tokens cost exactly 5× input tokens across the whole line, and Sonnet 5’s promo rate has a hard stop—modeling long-term unit economics on $2/$10 is a budget error waiting to happen.

Why Output Tokens Drive Your Claude API Bill (5× the Input Cost)

Forget the instinct to trim your system prompt; the real lever is response length. Every Claude model in this generation charges five times more per token that the API writes than per token it reads. A verbose output format will cost you far more than a long prompt ever will.

Take a typical Sonnet 5 request under intro pricing with 1 million input tokens and just half that much output:

const inputTokens  = 1_000_000;   // 1M tokens
const outputTokens =   500_000;   // 500k tokens

const inputPrice  = 2;    // $2 per million tokens (intro)
const outputPrice = 10;   // $10 per million tokens (intro)

const cost = (inputTokens  / 1e6) * inputPrice
            + (outputTokens / 1e6) * outputPrice;
// cost = $2 + $5 = $7

Even though the output is half the size, it’s responsible for over 70% of the $7 bill. When the intro rate expires and the same request hits $3/$15, the split becomes $3 + $7.50 = $10.50—still dominated by output.

If you’re building a Claude-backed feature, our AI development team spends as much time on response shaping and structured output schemas as on prompt engineering, precisely because that 5× ratio makes output length the budget’s loudest dial.

The Introductory Price Trap: Sonnet 5’s Promo Rate Expires

Anthropic’s Sonnet 5 introductory pricing—$2/M input, $10/M output—is current until 31 August 2026. On 1 September 2026, the standard rate of $3/$15 kicks in.

That’s a 50% increase on input and a 50% increase on output. Anyone building a product whose unit economics rely on the intro tier needs to bake that step-change into their runway model now. There’s no guarantee of an extension, and the standard rate is the number you’ll be paying for 99% of the product’s lifetime.

Beyond List Prices: Prompt Caching and Batch Lower Your Effective Rate

The table above shows the standard, on‑demand tier—the ceiling. In practice, two features can pull your effective Claude API cost well below the list numbers:

• Prompt caching: When you re-use a large system prompt or context across many calls, caching can slash the input price for the repeated portion.
• Batch processing: Submitting jobs asynchronously with longer latency windows often earns a substantial discount. The exact multiplier varies, but it’s enough to turn the list price from “budget assumption” to “worst-case scenario.”

The key: these are not footnotes. If your workload has any regularity, model your costs with caching and batch factored in, not the rack rate. The LLM API pricing calculator lets you toggle those levers on your own token volumes, so you can compare effective rates instead of guessing.

Model Your Own Usage—Don’t Trust a Worked Example

This article includes one code snippet; your own traffic patterns are the only ones that matter. Visit the LLM API pricing calculator, enter your expected monthly input and output tokens, and see a dated, per-model breakdown that you can screenshot and share with your team. If you’re ready to architect the feature on the right model from day one, start a project with us.