Task budgets

Task budgets let you tell Claude how many tokens it has for a full agentic loop, including thinking, tool calls, tool results, and output. The model sees a running countdown and uses it to prioritize work and finish gracefully as the budget is consumed.



When to use task budgets

Task budgets work best for agentic workflows where Claude makes multiple tool calls and decisions before finalizing its output to await the next human response. Use them when:

• You want Claude to self-regulate token spend on long-horizon tasks.
• You have a predictable per-task cost or latency ceiling to enforce.
• You want the model to finish gracefully (summarize findings, report progress) as it approaches the budget rather than cutting off mid-action.

Task budgets complement the effort parameter: effort controls how thoroughly Claude reasons about each step, while task budgets cap the total work Claude can do across an agentic loop.



Setting a task budget

Add task_budget to output_config and include the beta header:

client = anthropic.Anthropic()

with client.beta.messages.stream(
    model="claude-opus-5",
    max_tokens=128000,
    output_config={
        "effort": "high",
        "task_budget": {"type": "tokens", "total": 64000},
    },
    messages=[
        {"role": "user", "content": "Review the codebase and propose a refactor plan."}
    ],
    betas=["task-budgets-2026-03-13"],
) as stream:
    response = stream.get_final_message()

print(response.usage)

The task_budget object has three fields:

• type: always "tokens".
• total: the number of tokens Claude can spend across the agentic loop, including thinking, tool calls, tool results, and output.
• remaining (optional): the budget remainder carried over from a prior request. Defaults to total when omitted.



How the budget countdown works

Claude sees a budget-countdown marker injected server-side throughout the conversation. The marker shows how many tokens remain in the current agentic loop and updates as the model generates thinking, tool calls, and output, and as it processes tool results. Claude uses this signal to pace itself and finish gracefully as the budget is consumed.



Worked example: budget counting across turns

The task budget counts what Claude sees (thinking, tool calls and results, and text), not what's in your request payload. In an agentic loop your client resends the full conversation on every request, so the payload grows turn over turn, but the budget only decrements by the tokens Claude sees this turn.

Consider a loop with task_budget: {type: "tokens", total: 100000} and a single bash tool.

Turn 1. You send the initial request:

{
  "messages": [
    { "role": "user", "content": "Audit this repo for security issues and report findings." }
  ]
}

Claude thinks, then emits a tool call and stops with stop_reason: "tool_use":

{
  "role": "assistant",
  "content": [
    {
      "type": "thinking",
      "thinking": "I'll start by listing dependencies to look for known-vulnerable packages..."
    },
    {
      "type": "tool_use",
      "id": "toolu_01",
      "name": "bash",
      "input": { "command": "cat package.json && npm audit --json" }
    }
  ]
}

Suppose this assistant turn (thinking plus the tool call) totals 5,000 generated tokens. The countdown Claude saw during generation ended near remaining ≈ 95,000.

Turn 2. Your client runs the tool, then resends the full history with the tool result appended:

{
  "messages": [
    { "role": "user", "content": "Audit this repo for security issues and report findings." },
    {
      "role": "assistant",
      "content": [
        { "type": "thinking", "thinking": "I'll start by listing dependencies..." },
        {
          "type": "tool_use",
          "id": "toolu_01",
          "name": "bash",
          "input": { "command": "cat package.json && npm audit --json" }
        }
      ]
    },
    {
      "role": "user",
      "content": [
        {
          "type": "tool_result",
          "tool_use_id": "toolu_01",
          "content": "<2,800 tokens of npm audit output>"
        }
      ]
    }
  ]
}

The resent turn-1 user and assistant messages are not counted again, but the 2,800-token tool result is new content Claude sees this turn and counts against the budget. Claude spends another 4,000 tokens on thinking and a second tool call (grep -rn "eval(" src/). The countdown ends near remaining ≈ 88,200.

Turn 3. Full history resent again with the second tool result (1,200 tokens of grep output) appended. Claude writes a 6,000-token final findings report and stops with stop_reason: "end_turn". remaining ≈ 81,000.

Putting the three turns side by side makes the distinction between payload size and budget spend explicit:

Your client sent the turn-1 user message three times and the turn-1 assistant message twice, but each was counted once. The budget spent 19,000 of 100,000 tokens, even though the cumulative payload your client transmitted was larger and the prompt-cached input on turns 2 and 3 was larger still.



Carrying a budget across compaction with remaining

If your agentic loop compacts or rewrites context between requests (for example, by summarizing earlier turns), the server has no memory of how much budget was spent before compaction. Pass remaining on the next request so the countdown continues from where you left off rather than resetting to total:

# Tokens spent before compaction, tracked client-side
tokens_spent_so_far = 45000

output_config = {
    "effort": "high",
    "task_budget": {
        "type": "tokens",
        "total": 128000,
        "remaining": 128000 - tokens_spent_so_far,
    },
}

For loops that resend the full uncompacted history on every turn, omit remaining and let the server track the countdown.



Changing the budget mid-conversation

task_budget is a request-level setting. To change the budget partway through a task, for example to extend it when the user broadens the request, set a new task_budget in output_config on the next request. Keep the caching consequence in mind: the budget value participates in the rendered prompt, so a changed value does not match cache entries created under the old one (see Feature support below).



Task budgets are advisory, not enforced

Task budgets are a soft hint, not a hard cap. Claude may occasionally exceed the budget if it is in the middle of an action that would be more disruptive to interrupt than to finish. The enforced limit on total output tokens is still max_tokens, which truncates the response with stop_reason: "max_tokens" when reached.

For a hard cap on cost or latency, combine task budgets with a reasonable max_tokens value:

• Use task_budget to give Claude a target to pace against.
• Use max_tokens as the absolute ceiling that prevents runaway generation.

Because task_budget spans the full agentic loop (potentially many requests) while max_tokens caps each individual request, the two values are independent; one is not required to be at or below the other.



Choosing a budget

The right budget depends on how much work your agentic loop currently does. Rather than guessing, measure your existing token usage first and then tune from there.



Measure your current usage

Run a representative sample of tasks without task_budget set and record the total tokens Claude spends per task. For an agentic loop, sum usage.output_tokens across every request in the loop, plus the tokens of the tool results you append between requests:

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-5",
    max_tokens=4096,
    messages=[
        {"role": "user", "content": "Review the codebase and propose a refactor plan."}
    ],
)

# Sum output_tokens (text + thinking + tool calls) across every request in your loop.
print(response.usage.output_tokens)

Run this across a representative set of tasks and record the distribution. Start with the p99 of your per-task token spend to understand how providing the model with a task budget might modify the model's behavior, then test up or down as needed.

The minimum accepted task_budget.total is model-specific; on every model that currently supports task budgets (see Feature support) it is 20,000 tokens, and values below the minimum return a 400 error.



Interaction with other parameters

• max_tokens: Orthogonal to task budgets. max_tokens is a hard per-request cap on generated tokens, while task_budget is an advisory cap across the full agentic loop (potentially spanning many requests). At xhigh or max effort, set max_tokens to at least 64k to give Claude room to think and act on each request.
• Effort: Effort controls how deeply Claude reasons per step. Task budgets control how much total work Claude does across an agentic loop. The two are complementary: effort tunes depth, task budgets tune breadth.
• Adaptive thinking: Task budgets include thinking tokens in the count, so adaptive thinking naturally scales down as the budget depletes.
• Prompt caching: The budget-countdown marker is injected server-side per turn, so it does not match across requests. If your client decrements task_budget.remaining on each follow-up request, the changed value invalidates any cache prefix that contains it. To preserve caching, set the budget once on the initial request and let the model self-regulate against the server-side countdown rather than mutating the budget client-side.



Feature support

Task budgets are not supported on Claude Code or Cowork surfaces. Use task budgets directly through the Messages API on a supported model.



Next steps