Claude Opus 5.5 Effort Levels: Why Medium Is the New Default

What the effort parameter does

Effort tells Claude how thorough to be. It sets how deeply the model thinks, and it also shapes the rest of its behavior: at lower effort it makes fewer, more consolidated tool calls, writes less preamble and gives shorter confirmations. Because thinking tokens bill as output, effort moves cost and latency more than any other setting.

response = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=16000,
    output_config={"effort": "medium"},  # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Review this pull request for bugs."}],
)

Effort sits inside output_config, not at the top level of the request. No beta header is needed.

The five effort levels

At low, how often Opus 5.5 skips thinking altogether depends on the prompt. Simple questions often get a near-instant answer, while complex ones still get some reasoning.

Why medium is the new default

Effort names aren’t a fixed amount of thinking. They are relative to each model. Opus 5.5 does more with each level than Opus 5 did, so Anthropic moved the default down a step. In its testing, Opus 5.5 at medium exceeded Opus 5 at high on coding and knowledge-work evaluations.

Customer results from the launch show how far down the scale the quality holds:

• Deloitte: Opus 5.5 at low caught 72% of known bugs. Opus 5 at high caught 56%, with more false alarms.
• Walleye Capital: solved its evaluation task at the lowest setting, and caught an off-by-one error in the task's own instructions.
• Rogo: beat Opus 5 at high effort using 60% fewer output tokens.
• Factory: matched Opus 5 at high effort with 20 to 25% fewer output tokens.

The biggest gap is visual work. Anthropic reports that Opus 5.5 at low effort read charts more accurately than Opus 5 at its highest setting, using about a tenth of the output tokens per chart.

How to choose a level

The right level depends on the workload, so tune it per route rather than once for the whole app. A short sweep gives you the answer:

1. Take 50 to 100 real requests from the route you are tuning.
2. Run them at low, medium and high. Add xhigh only for hard agentic coding.
3. Grade the outputs with the same checks you use in production: tests, schemas or a reviewer.
4. Work out the cost per completed task at each level, not the cost per request.
5. Pick the lowest level that holds your quality bar.

If Opus 5.5 still fails at high or xhigh, the next step may be a different model rather than more effort. See Opus 5.5 vs Fable 5.1 for when to escalate.

Changing effort mid-conversation

Agent loops mix hard steps and routine ones. Changing the top-level effort between requests works, but it invalidates the prompt cache. The per-message form avoids that: an empty system message that carries only the new effort level, under the mid-conversation-output-config-2026-07-01 beta.

POST /v1/messages
anthropic-beta: mid-conversation-output-config-2026-07-01

{
  "model": "claude-opus-5-5",
  "max_tokens": 16000,
  "output_config": {"effort": "high"},
  "messages": [
    {"role": "user", "content": "Plan the database migration."},
    {"role": "assistant", "content": "Here's the plan: ..."},
    {"role": "system", "content": [], "output_config": {"effort": "low"}},
    {"role": "user", "content": "Now rename the config file."}
  ]
}

The new level applies from the next user turn until another system message changes it. Lowering effort this way is reliable. Raising it works best in large jumps, such as low to xhigh.

Common mistakes

• Leaving effort unset. You get medium, which may be more or less than you want. Set it on every route.
• Prompting instead of lowering effort. “Be brief” or “don’t overthink” is less reliable than a lower level. Lower effort first, then prompt if needed.
• A max_tokens too small for the thinking. Thinking counts toward the limit. A limit sized for Opus 5 with thinking off can cut replies short. Anthropic suggests 64K for long agentic coding turns.
• Using xhigh or max by default. Reserve them for work where you measured a gain, and plan for longer turns with streaming and timeouts.

To see what effort means for your monthly bill, read the Opus 5.5 pricing guide. For the other API changes, see the Claude Opus 5.5 migration guide.