Home / Pricing guides / AI Model API / Claude API
Pricing guide, AI Model API
Input and output rates by model tier, the Batch and caching discounts worth engineering around, and why the model name in your code matters more than any negotiation.
The short answer
Claude API pricing in the sources used here runs from USD 1 per million input tokens and USD 5 output on Haiku 4.5 to USD 10 and USD 50 on Fable 5, with Sonnet 4.6 at USD 3 and USD 15. At 100 million input and 20 million output tokens a month, Sonnet 4.6 costs USD 7,200 a year, or USD 3,600 through the Batch API.
Written by Procurement VMS Editorial Team · Reviewed by Michael, Procurement Advisor Expert · Price data dated · 11 minute read
The Claude API has no plans, no seats and no renewal date, which makes it easy to read and hard to budget. You pay per million tokens, with input and output priced separately, and the spread across models is wide enough that the model name in your code matters more than anything a procurement team could negotiate.
This guide prices four model tiers as listed in two dated 2026 sources, and explains the two discounts worth engineering around: 50% off through the Batch API and 90% off on cache reads.
Prices below were recorded on 17 October 2026. The headline figure is the lowest price you can get for that plan. Check the official page before you buy, because vendors change prices without notice.
USD1 input, USD 5 output
per million tokens
Best for: High-volume classification, extraction and routing where cost per call decides the business case
USD3 input, USD 15 output
per million tokens
Best for: The default production workhorse for most applications
USD5 input, USD 25 output
per million tokens
Best for: Hard reasoning and agentic work where a better answer is worth five times the input cost
USD10 input, USD 50 output
per million tokens
Best for: The frontier tier, used selectively rather than as a default
How billing works
There are no seats and no plans: you are billed for tokens, so the total depends on your usage and nobody can quote an annual figure without a volume estimate. Three rules do most of the work on the invoice. Output tokens cost five times input tokens on every model, so a chatty application costs far more than a token count suggests. The Batch API takes a flat 50% off both input and output for asynchronous jobs, and it stacks with caching. Prompt caching charges more than base input to write, 1.25 times for a 5-minute cache and 2 times for a 1-hour cache, but cache reads cost a tenth of base input, so caching pays only once a cached prefix is read several times. Fast Mode is priced separately at a multiple of the standard rate and is not available through the Batch API.
The tier trap
The expensive decision is the default model name sitting in your code and not a plan you sign. Moving a production workload from Sonnet 4.6 to Opus multiplies input and output cost by 1.67 for the same traffic, and Fable 5 is ten times Haiku on input. Fast Mode compounds it, because it is priced at a multiple of the standard rate and cannot use the 50% Batch API discount.
Read this first. Medium confidence. Anthropic's official pricing page could not be fetched for this guide. Token rates come from two independent 2026 guides that cite Anthropic's pricing page and agree on the figures used: one dated 5 April 2026 and one reviewed 17 June 2026. The Fable 5 rate comes from the second guide only. Model names and versions move quickly, newer versions have been released since those sources, and the sources disagree on which Opus version was current, so confirm the current model line-up and rates on Anthropic's pricing page before you budget. Fast Mode surcharges are described without a number for the same reason. The OpenAI comparison rate comes from BenchLM, last synced 16 October 2026. This is a vendor product page about Claude written by Procurement VMS and it states published rates without preference.
Move the slider to see what each plan costs at your size. The figures use the list prices on this page and leave out tax, add-ons and any discount you negotiate.
One workload priced three ways shows how model choice and batching change the same traffic.
| Scenario | Cost per year | Against the headline |
|---|---|---|
| Sonnet 4.6 at 100M input and 20M output tokens a month, standard API | USD7,200 | Headline price |
| Same volume moved to Opus, standard API | USD12,000 | +USD 4,800 |
| Same volume on Sonnet 4.6 run through the Batch API | USD3,600 | -USD 3,600 |
Compare the plans side by side. Where a cell says Check page, the vendor's pricing page was not specific enough for us to fill it in.
| Feature | Haiku 4.5 | Sonnet 4.6 | Opus | Fable 5 |
|---|---|---|---|---|
| Input per million tokens | USD 1 | USD 3 | USD 5 | USD 10 |
| Output per million tokens | USD 5 | USD 15 | USD 25 | USD 50 |
| Batch API input per million tokens | USD 0.50 | USD 1.50 | USD 2.50 | USD 5 |
| Batch API output per million tokens | USD 2.50 | USD 7.50 | USD 12.50 | USD 25 |
| Cache read, a tenth of base input | USD 0.10 | USD 0.30 | USD 0.50 | USD 1 |
| 1-hour cache write, twice base input | USD 2 | USD 6 | USD 10 | USD 20 |
| Context window | Around 200K | 1M at a flat rate | 1M at a flat rate | 1M at a flat rate |
| Fast Mode available at a surcharge | Not published | Not published | Check page | Not published |
Pick it for classification, extraction or routing at volume, where cost per call decides the business case.
Pick it as the production default for most applications, with a 1M context window at a flat rate.
Pick it for hard reasoning and agentic work where answer quality is worth five times the Haiku input rate.
Pick it selectively for frontier work, and not as a default, because it is ten times Haiku on input.
Our view
Token pricing is easy to model, which makes the Claude API good value for teams that will do the modeling. The spread from USD 1 to USD 10 per million input tokens means model selection, and not negotiation, is where the money is, and the 50% Batch discount and 90% cache-read discount are large enough to change a build decision. It is a poor fit for a finance team that wants a fixed annual number.
You may be overpaying if
You default everything to Opus or Fable 5 when Sonnet or Haiku would pass, you run asynchronous jobs through the standard API and not the Batch API, or you cache prefixes that are read once.
Haiku 4.5 is USD 1 per million input tokens against USD 5 on Opus. Classify, extract and route on Haiku, and reserve Opus for work that needs it.
The Batch API is a flat 50% off input and output on every model and it stacks with caching. On 100M input and 20M output tokens a month on Sonnet 4.6 that is USD 3,600 a year.
Cache reads cost a tenth of base input, USD 0.30 against USD 3 on Sonnet 4.6. The write costs 1.25x or 2x, so only cache a prefix you will read repeatedly.
Output is 5x input on every model. Capping response length and tightening output formats moves the bill more than trimming prompts.
Fast Mode carries a multiple of the standard rate and cannot use the Batch API. Make it an explicit per-feature decision and not a global setting.
This bill has no renewal date to catch it, so the only control is a monitored threshold that someone owns and reviews monthly.
Renewal email you can copy
Fill in the brackets and send it about 90 days before the renewal date. It works best next to your renewal tracker.
Subject: Claude API renewal on [renewal date] Hi [account manager], Our Claude API [plan] subscription renews on [renewal date]. Over the last 90 days, [active million input tokens per month] of the [paid million input tokens per month] we pay for used it. We would like to renew with [target million input tokens per month] on a [term] term. Before we sign, please send: 1. The current list price and the price you are offering us, line by line. 2. The cap on any price increase at the next renewal. 3. The notice period if we want to cut million input tokens per month or move to another plan. We are comparing this against [alternative] and plan to decide by [decision date]. If you can offer [target discount] percent off list, or a lower price for a longer term, we are ready to talk this week. Thank you, [name] [company]
Free PDF and Excel
A one page summary of every plan, plus an Excel model that adds up your own seat count.
If this vendor does not fit, these are the closest options and what they charge.
| Alternative | Starting price | Pick it when |
|---|---|---|
| OpenAI API | From USD 0.10 per million input tokens and USD 0.50 output (GPT-6 Luna) | Your cheapest tier is doing high-volume simple work Read our OpenAI API pricing guide |
| Claude Team seats | USD20 per seat per month (annual billing) | Your people need an assistant rather than your product needing an API Read our Claude pricing guide |
| Claude Enterprise | USD20 per seat per month plus usage at API rates (annual) | You want the same token rates with SCIM, audit logs and spend limits around them Read our Claude pricing guide |
A usage bill sends no renewal notice. Give the spend a named owner and a monthly threshold review.
Bottom line. Default to a mid-tier model, push volume work to the smallest tier, put every asynchronous job through the Batch API, and set a spend threshold with a named owner. Confirm current model names and rates before you budget.
In the sources used here, Haiku 4.5 is USD 1 per million input tokens and USD 5 per million output, Sonnet 4.6 is USD 3 and USD 15, Opus is USD 5 and USD 25, and Fable 5 is USD 10 and USD 50. Output is charged at five times input on every model, which is where most budgets go wrong. Confirm current models and rates on Anthropic's pricing page.
No. The API is pay as you go with no published free token allowance and no trial. If you want a no-cost way to evaluate Claude, that is the consumer Free plan on claude.com and not the API, and it is not a development environment.
There is no enterprise API price list, because the rate is the same per token whatever your size and the total is set by volume. Claude Enterprise seats are USD 20 per seat per month on annual billing plus usage billed at these same API rates, with spend limits and audit logs around it.
Yes, but not by negotiating. The Batch API is a flat 50% off input and output on every model for asynchronous jobs, and cache reads cost a tenth of base input. The two stack, so a cached, batched workload can run well under half the headline rate.
There is no published annual or committed-use discount, so there is nothing to prepay for. Your structural levers are model choice, the 50% Batch API discount and prompt caching, all of which are engineering decisions and not procurement ones.
On headline rates the OpenAI API goes lower at the small end, from USD 0.10 per million input tokens and USD 0.50 output on its smallest model, per BenchLM. Compare at the model you would actually use and not at the cheapest tier, because capability per token decides the real cost.
We start from the vendor pricing page and note the date. Where a figure comes from a third party instead, we say so on the page. Procurement VMS is a ratings and comparison site, not a software vendor, and nobody paid for a place in this guide. Spot a number that has changed? Tell us on the contact page and we will check it.
You get a one page PDF and an Excel model with your seat count, the plans above and a renewal checklist.
Choosing software instead of renewing? See the ranked procurement software list.
