Claude Agent SDK usage can draw from plan limits, from the monthly API credits added to Max and Team plans on October 7, or from pay-per-token billing. This article sets out every price, the credit amounts, cache savings and a worked cost for one agent run.
The Claude Agent SDK has no price of its own. The bill comes from the model behind it, and as of October 8, 2026 that bill can reach you in three different ways: through your subscription limits, through the new monthly API credits on Max and Team plans, or through plain pay-per-token billing. Pick the wrong route and you either hit a wall in the middle of a run or pay more than you had to. This article sets out each option with the numbers from Anthropic's own pricing and help pages, then works out what a single agent run costs on three different models.
The Agent SDK is a Python and TypeScript library that puts Claude Code's agent loop, built-in tools, sessions and hooks inside your own program. Anthropic's documentation describes it as a library that runs the Claude Code binary in a process you operate. Its pricing page lists no SDK fee, so what you pay for is model usage: input tokens, output tokens and cache activity.
Where those tokens get charged depends on how the SDK signs in:
| Route | Billed in | Best for | Watch out for |
|---|---|---|---|
| Subscription limits | Plan allowance | Personal scripts and nightly jobs | Sharing one pool with your chats |
| Monthly credits | Dollars that expire | Prototypes and side projects | No rollover at the end of the cycle |
| Pay-per-token | Dollars you buy | Products with real users | No cap unless you set one |
💡 Cloud providers are a fourth route. The SDK can run on Amazon Bedrock, Google Cloud or Microsoft Foundry. Those bills come from the cloud provider, and the Max and Team credits do not apply there.
Search results are full of conflicting claims because the rules moved three times this year:
claude -p, Claude Code GitHub Actions and third-party apps built on the SDK would move to a separate monthly credit billed at full API rates: $20 on Pro, $100 on Max 5x and $200 on Max 20x.If an article says Pro subscribers receive a $20 SDK credit, it is describing the plan that was paused.
Anthropic's help article for the SDK says usage that goes through your Claude login keeps counting against your subscription limits, the same as claude -p. That is the right route for personal automation: a script on your laptop, a nightly job on your own machine. A subscription costs $20 a month for Pro, $100 for Max 5x and $200 for Max 20x, and you pay nothing per token.
The catch is sharing. The SDK draws from the same pool as every chat and every interactive coding session you run, so a long agent loop can leave you short of allowance for the rest of your day.
On October 7, 2026, Anthropic added monthly API credits to Max and Team plans. They apply to the Claude API, Claude Managed Agents, the Claude Agent SDK and the Console playground.
| Plan | Monthly credit |
|---|---|
| Max 5x | $100 |
| Max 20x | $200 |
| Team, Standard seat | $20 per seat |
| Team, Premium seat | $100 per seat |
| Free, Pro and Enterprise | Not eligible |
A credit is spent at the API prices in the next section, so $100 means $100 of tokens. Interactive Claude Code and extra usage inside the Claude apps are excluded, and the credits leave your plan limits unchanged.
Credits are not automatic. Follow these steps:
A few rules decide how far the money goes:
Your credit balance is too low to access the Anthropic API.These are list prices per million tokens from Anthropic's pricing page, checked on October 8, 2026.
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $0.25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
| Claude Sonnet 5 | $2 | $10 | $0.20 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 | $0.50 | $0.01 |
| Claude Haiku 4.5 | $1 | $5 | $0.10 |
Claude Haiku 5.5 costs more when a prompt passes 100,000 tokens: $0.50 for input and $2.50 for output.
💡 Claude Sonnet 5 launched with $2 and $10 as introductory pricing through August 31, 2026. Anthropic's page now says that is the standard price and the planned September 1 rise to $3 and $15 will not happen. Any page still listing that rise is out of date.
Models from version 4.7 onward use a newer tokenizer that produces roughly 30% more tokens for the same text than the 4.6 generation and earlier. Compare the cost of a finished task, not only the price per token.
The SDK uses prompt caching on its own. A 5-minute cache write costs 1.25 times the base input price, a 1-hour write costs 2 times, and a cache read costs 0.1 times. Opus 5.5 and Sonnet 5.5 read at 0.05 times, and Fable 5.1 at 0.025 times. An agent loop resends a long, stable prefix on every step, so most of its input tokens end up as cheap reads.
With an API credential the cache lives for 5 minutes by default. If your runs sit further apart than that, set the ENABLE_PROMPT_CACHING_1H environment variable. The 1-hour option costs more to write and pays off when later runs read the same prefix.
Tokens are not the whole bill. Check these before you budget:
inference_geo: "us" on Claude 4.6 and later models.Picture a bug-fix run: the agent reads a few files, edits code, runs the tests and repeats for about 30 steps. Assume 600,000 input tokens across those steps, made up of 20,000 uncached tokens, 40,000 written to the 5-minute cache and 540,000 read back from it, plus 30,000 output tokens. These are illustrative numbers, not a measurement, but they have the shape of a cached agent loop.
On Sonnet 5.5 the arithmetic is:
That adds up to $0.494. Run the same shape on three models and one $100 credit buys very different amounts of work:
| Model | Cost per run | Runs from a $100 credit |
|---|---|---|
| Claude Haiku 5.5 | about $0.03 | more than 3,600 |
| Claude Sonnet 5.5 | about $0.49 | about 200 |
| Claude Opus 5.5 | about $0.99 | about 100 |
Two lessons stand out. First, caching saves roughly two thirds: the same Sonnet 5.5 run with no cache would cost $1.50. Second, output is the expensive part. Those 30,000 output tokens are about 5% of the volume and 61% of the cost. A Max 20x credit simply doubles every number in the last column.
Every query() call ends with a result message that carries total_cost_usd, the estimated cost of that call. The modelUsage field breaks it down per model, including cache reads and writes. Three details trip people up:
total_cost_usd and modelUsage include subagent work, but usage leaves it out, so usage undercounts as soon as subagents run.output_tokens on each step is a placeholder. Read the real output count from the result message.💡
total_cost_usdis a client-side estimate built from a price table bundled with the SDK. Anthropic says to use the Usage and Cost API or the Usage page in the Console for authoritative billing, and never to bill your own customers from the estimate.
import { query } from "@anthropic-ai/claude-agent-sdk";
for await (const message of query({
prompt: "Fix the failing test in src/auth.ts",
options: { maxBudgetUsd: 2 },
})) {
if (message.type === "result") {
console.log(message.subtype, message.total_cost_usd);
}
}
The maxBudgetUsd option in TypeScript, or max_budget_usd in Python, stops a call once its own spend crosses the limit. It counts only that call: spend restored from a resumed session does not count against it, and a /clear starts the budget over. The result then arrives with the subtype error_max_budget_usd.
For a ceiling across a whole project, give it its own Console workspace and set that workspace's spend limit. Every credential and workspace in a linked organization draws from the same credit balance, so a workspace limit stops one noisy job from eating the month's allowance.
Team credits are $20 per Standard seat and $100 per Premium seat, pooled into one balance capped at $500 a month. A team with three Standard seats and two Premium seats receives $260 a month, and adding a Premium seat lifts the next month's credit to $360. The amount is set by the seats on the plan at the start of each billing month. Linking an organization also moves it to at least the Start usage tier, or the Build tier when the monthly credit is above $200.
If you ship a product, plan on API billing. Anthropic does not allow third-party developers to offer claude.ai login or its rate limits in their products, including agents built on the SDK, unless Anthropic approved it in advance. The monthly credits are meant for building and running your own applications and agents, so keep customer traffic on pay-per-token billing, where you control spend limits per workspace.
Claude Managed Agents is the hosted alternative: Anthropic runs the agent loop in a cloud sandbox it manages. It bills the same token rates plus $0.08 per session-hour, counted only while the session is running. Idle time does not count. A one-hour session on Sonnet 5.5 with 50,000 input tokens and 15,000 output tokens costs $0.10 + $0.15 + $0.08, or $0.33. With the SDK there is no runtime line, but the machine it runs on is yours to pay for. The Max and Team credits work on both.
Before you spend credits tuning an agent, draft and test its system prompt in a chat model. Claude Sonnet 5 is on PicassoIA, built for multi-step coding and tool-use tasks, which makes it a handy place to try a prompt before wiring it into the SDK.
💡 The effort ladder also moves your SDK bill. Higher effort means more output tokens, and output is the costly side of every price table above.
To compare how a lighter or heavier model answers the same prompt, Claude Haiku 4.5, Claude Opus 4.7 and Claude Fable 5 sit in the same collection.
Every photo in this article came from P Image, a text-to-image model on PicassoIA that returns a finished picture in under a second. It accepts 16:9, 1:1, 9:16 and custom sizes, offers prompt upsampling for short ideas, and takes a seed so a result you like can be reproduced exactly.
Try it now. Describe one scene in a single sentence, such as a desk, a window and late afternoon light. Run it, change a single detail, and run it again. Ten tries takes a few minutes, and by the end you will have a feel for how camera, light and texture words shape a photo.
PicassoIA also offers 91 text-to-image models and 87 text-to-video models, plus music, speech and language models, all in the full model list. Pick one, type a prompt and make something of your own today.
