Thinking Token Budgets and Context Window Impact for Databricks Agents

1.Thinking spends the output budget

Reasoning tokens are not free extras on top of the answer. They come out of the same output limit. The documented extended-thinking request for databricks-claude-sonnet-4-5 gives thinking a budget of 10,240 tokens inside an overall max_tokens of 20,480:

response = client.chat.completions.create(
    model="system.ai.claude-sonnet-4-5",
    messages=[{"role": "user", "content": "Why is the sky blue?"}],
    max_tokens=20480,
    extra_body={
        "thinking": {
            "type": "enabled",
            "budget_tokens": 10240
        }
    }
)

Two hard rules come with this. First, budget_tokens must be less than max_tokens. A thinking budget at or above the output limit is invalid, so the predict above describes a request you need to fix. Second, bigger budgets give diminishing and less predictable returns: "Higher budgets can improve quality for complex tasks, but usage above 32K may vary." The Responses API states the same accounting directly. Its max_output_tokens is an upper bound on generated tokens "including visible output tokens and reasoning tokens". Every token you give to thinking is a token the final answer can no longer use.

Sources12

2.Reasoning tokens in a multi-turn context window

In an agent loop, the question changes from "how much does this call cost?" to "what does this call leave behind in the context?" For databricks-claude-sonnet-4-5 the documentation answers it ("This section is specific to the databricks-claude-sonnet-4-5 model."): only the reasoning blocks associated with the last assistant turn or tool-use session are visible to the model and counted as input tokens. Thinking from earlier turns does not keep piling up as input. The most recent turn's reasoning does, so a large extended-thinking budget on the latest turn makes the next request's input larger.

You can also choose not to send that reasoning back. If the next step does not need the model to reason over its previous steps, you can omit the reasoning block entirely. That keeps those tokens out of the input context. Don't assume other models behave the same way. The documentation scopes this rule to Sonnet 4.5.

Sources1

3.Choosing and justifying a mode for an agent task

A sound justification has three parts: what the task needs, what setting delivers it, and what that setting costs in tokens and context. Hybrid Sonnet models are aimed at work that needs "a balance of practical throughput and advanced thinking such as customer-facing agents". In practice that means fast replies by default and thinking only on the hard steps. Claude Fable 5.1, built for long-running work, "supports per-turn effort controls and thinking updates between tool calls for sustained, complex workflows", so effort can change from step to step instead of being fixed for the whole agent.

Sources31

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.budget_tokens is a separate allowance on top of max_tokens, so you can set it as high as you like.Why is that wrong?

   Thinking uses up the output allowance. budget_tokens must be strictly less than max_tokens, and budgets above 32K behave less predictably.

   Covered in Thinking spends the output budget

2. 2.With extended thinking, reasoning from every earlier turn piles up in the input context of a multi-turn agent.Why is that wrong?

   For databricks-claude-sonnet-4-5, only the last assistant turn's or tool-use session's reasoning counts as input, and you can leave even that out.

   Covered in Reasoning tokens in a multi-turn context window