D1 · Agentic Architecture › Cluster 7 · Performance

7.2 · Adaptive Thinking & Effort as Knobs

Two dials that trade cost for quality

CONCEPT Two independent knobs

Current models expose two separate controls that shape how much compute a request spends. Adaptive thinking governs internal reasoning; effort governs how consolidated the visible output and tool-calling behavior is. Both trade cost/latency against thoroughness.

KnobParameterWhat it controls
Adaptive thinking thinking={"type":"adaptive"} Claude decides the reasoning depth and interleaves thinking between tool calls automatically — no manual budget.
Effort output_config={"effort": ...} How hard the model works on the visible response: "low" · "medium" · "high" (default) · "max". Some models add "xhigh".
⛔ The old fixed budget_tokens under thinking is removed on current models. Sending it returns a 400. Use {"type":"adaptive"} instead — Claude picks the depth for you.

CONCEPT How effort trades cost vs. quality

Lower effort tells the model to be terser and to consolidate tool calls — fewer, bigger steps and shorter output, so fewer tokens and lower latency. Higher effort spends more compute for more thorough exploration and output.

EffortBehaviorReach for it when…
"low"Fewest, most-consolidated tool calls; terse output Simple tasks; cost/latency-sensitive high-volume paths
"medium"Balanced Everyday production work
"high" (default)Thorough exploration & output Hard tasks; when quality matters more than a few tokens
"max"Maximum effort The hardest reasoning where you want every bit of headroom

DIAGRAM Where the knobs sit

flowchart LR R([Request]) --> T{"thinking:
adaptive?"} T -->|yes| TD["🟣 Claude picks reasoning depth
interleaved between tool calls"] T -->|no| TP["🔵 no interleaved thinking"] TD --> E{"output_config
effort"} TP --> E E -->|low| L["🟢 fewer tool calls · terser
cheaper · faster"] E -->|high / max| H["🟠 thorough · more tokens
higher quality · costlier"] classDef think fill:#e2d4f0,stroke:#764ba2,color:#3d1a5b; classDef none fill:#d1ecf1,stroke:#17a2b8,color:#0c5460; classDef cheap fill:#d4edda,stroke:#28a745,color:#155724; classDef dear fill:#ffe8cc,stroke:#d9480f,color:#7a2e08; class TD think; class TP none; class L cheap; class H dear;

CODE Adaptive thinking + effort together

The two knobs compose. Here a hard agentic task uses adaptive thinking with "high" effort; a cheap fan-out worker turns thinking off and drops to "low".

# Hard task: let Claude reason as deep as it needs, work thoroughly.
resp = client.messages.create(
    model="claude-opus-4-8",
    max_tokens=8192,
    thinking={"type": "adaptive"},      # Claude decides depth; interleaves
    output_config={"effort": "high"},   # default; thorough
    tools=TOOLS,
    messages=messages,
)

# Cheap, simple sub-task: no thinking, low effort → fewer calls, terser.
worker = client.messages.create(
    model="claude-haiku-4-5",
    max_tokens=512,
    output_config={"effort": "low"},
    messages=[{"role": "user", "content": f"Extract the total from: {line}"}],
)

# ⛔ WRONG on current models — budget_tokens is removed → HTTP 400:
# thinking={"type": "enabled", "budget_tokens": 4000}

TRADE-OFFS Turning the dials

✅ Lower effort / no thinking

  • Fewer, consolidated tool calls → lower latency
  • Terser output → fewer output tokens (the pricey ones)
  • Great for simple, high-volume, parallel sub-tasks

⚠️ Higher effort / adaptive thinking

  • More tokens & latency per request
  • But better on ambiguous, multi-step reasoning
  • Adaptive thinking removes the guesswork of a manual budget
Default is effort:"high". Down-tier effort on roles that measurably tolerate it, exactly like you down-tier the model (7.1).

EXAM TRAPS

Trap: Setting thinking={"budget_tokens": N} to "control cost." → That param is removed on current models and returns a 400. Use {"type":"adaptive"} and let Claude pick depth.
Trap: Confusing the two knobs. → Thinking = reasoning depth (adaptive); effort = how consolidated the output/tool-calling is. They're separate params.
Trap: Assuming a numeric effort. → Effort is a named level (low/medium/high/max), not a token count. Default is "high".

PRACTICE

Q1 · You want to cut latency on a high-volume routing step without changing the model. Which knob?

Set output_config={"effort":"low"}. Low effort yields fewer, more-consolidated tool calls and terser output → faster and cheaper, no model change needed.

Q2 · A request that previously set budget_tokens now 400s. Fix it.

Remove budget_tokens and use thinking={"type":"adaptive"}. The fixed budget was removed on current models; adaptive lets Claude choose depth and interleave thinking between tool calls.

Q3 · How does adaptive thinking behave across a multi-tool agent turn?

Claude decides the depth and interleaves thinking between tool calls automatically — you don't schedule it. Combine with an appropriate effort level to tune overall cost vs. quality.