Key takeaways
- Medium is the default effort for Claude Opus 5.5 in the API, the Claude apps and Claude Code, one level below Opus 5's default of high.
- On Anthropic's SWE-bench Pro subset, high solved 95.3% of tasks at $0.29 each, against 92.8% at $0.21 for medium.
- Artificial Analysis scores Opus 5.5 at 52 on high and 48 on medium, with high using about 32% more output tokens.
On Claude Opus 5.5, medium vs high comes down to this: run medium for most work and switch to high only when a task needs more checking. Medium is the model's default in the API, the Claude apps and Claude Code. Anthropic's own coding test puts high about 2.5 points ahead of medium at about 38% more cost per task. The effort docs describe high as the level for "complex reasoning, difficult coding problems, agentic tasks".
The question got more common this week. On November 12, 2026, Anthropic published a guide to using Opus 5.5 for everyday work, and Opus 5.5 is now the default model on the Pro, Max, Ultra and Team plans. The guide gives two pieces of advice: drop "think hard" from your prompts, and start at Medium. The Decoder covered the guide on November 14.
This comparison is based on Anthropic's documentation, its published test results and independent numbers from Artificial Analysis. It was not tested hands-on.
What effort does on Claude Opus 5.5
Effort controls how many tokens Claude spends on a reply. That includes its thinking, the text you read and any tool calls it makes. On Opus 5.5 effort is the main dial, because thinking is always on. The model decides how much to think based on the request and the effort level. An API request that tries to switch thinking off (thinking: {"type": "disabled"}) gets a 400 error at every level, according to What's new in Claude Opus 5.5.
There are five levels: low, medium, high, xhigh (shown as "Extra high" in the apps) and max. Effort doesn't change the price per token. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens at every level, as of November 14, 2026. A higher level costs more because the model writes more tokens, and it takes longer for the same reason.
Anthropic says effort is "a behavioral signal, not a strict token budget." At a lower level the model still thinks on hard problems. It just thinks less than it would at a higher level. Lower effort also means fewer tool calls, shorter confirmations and less explaining before it acts.
Claude Opus 5.5 medium vs high, side by side
| Medium | High | |
|---|---|---|
| Default? | Yes, in the API, the Claude apps and Claude Code | No (high is the default on most other Claude models) |
| What Anthropic says it's for | "Agentic tasks that require a balance of speed, cost, and performance" | "Complex reasoning, difficult coding problems, agentic tasks" |
| Anthropic's SWE-bench Pro subset | 92.8% of tasks solved, $0.21 per task | 95.3% of tasks solved, $0.29 per task |
| Artificial Analysis Intelligence Index | 48 | 52 |
| Output tokens to run that index | 42.5 million | 56.1 million |
| Time per index task | 6.9 minutes | 9.5 minutes |
| Time to first answer token | 17.6 seconds | 27.7 seconds |
| Output speed | 105 tokens per second | 106 tokens per second |
| Price per million tokens | $4 input, $20 output | $4 input, $20 output |
The SWE-bench Pro rows come from Anthropic's page on optimizing for cost and intelligence. The other numbers come from Artificial Analysis's Claude Opus 5.5 page, read on November 14, 2026. Output speed is almost the same at both levels. High takes longer because it writes about 32% more tokens, not because it writes them more slowly.
How much better high is, and what it costs
In Anthropic's coding test, going from medium to high raises the share of solved tasks from 92.8% to 95.3%. The cost per task goes from $0.21 to $0.29, about 38% more. On the Artificial Analysis Intelligence Index, high scores 4 points higher (52 vs 48). It uses about 32% more output tokens, and each task takes about 38% longer.
The two levels are close, and that is the main finding. The big jumps are at the ends of the scale. In Anthropic's test, low solved 87.4% of tasks for $0.11. Xhigh solved 96.7% for $0.74, which is about 2.5 times the cost of high. Artificial Analysis shows the same pattern. Low scores 35 on its index and max scores 59. Max takes 25.4 minutes per task and waits about 400 seconds before its first answer token.
Anthropic also changed what each level name means on this model. Its prompting guide for Claude Opus 5.5 says Opus 5.5 at medium "matches or exceeds" Claude Opus 5 at high on its coding and knowledge-work tests. On a subset of SWE-bench Pro, Opus 5 at high solved 91.9% of tasks and Opus 5.5 at medium solved 92.8%. Anthropic puts Opus 5.5's cost per solved task 78% lower. Factory, a coding-agent company quoted in Anthropic's guide, reported the same match with 20 to 25% fewer output tokens. That is a customer's claim, published by Anthropic.
Anyone moving from Opus 5 should know one more thing. At the same level name, Opus 5.5 tends to think more per turn than Opus 5, most of all at xhigh and max. If you keep the xhigh you used on Opus 5, expect longer turns and bigger bills.
When to pick medium and when to pick high
Anthropic gives the same advice everywhere: start at medium, measure, and raise effort only where it helps. In practice:
- Use medium for everyday chat, writing, analysis, agentic coding on well-defined tasks, and anything where you read each result yourself. It is the default Anthropic picked for this model, and its tests show it matching Opus 5 at high.
- Use high where checking matters or edge cases are likely. The Claude Code docs give "fixing a bug in an existing codebase" as the example. The prompting guide also says the model makes better use of image tools at higher effort, which helps with dense technical drawings.
- Use low for quick back-and-forth, subagents and simple high-volume requests. Anthropic notes that low is much faster on some tests. On the DRACO test a typical task took 84% less time than at medium, but scored 13.3 points lower.
- Keep xhigh and max for work where you have measured a gain. Anthropic's prompting guide says to "reserve
xhighandmaxfor work where you've measured a quality gain."
If medium falls short on one kind of task, raise effort for that task only. Don't add "think hard" to your prompts instead. Anthropic says that when it removed such lines in a chat product, replies started sooner, "with no clear decline in the quality of the reply." Lowering or raising effort changes how much the model thinks more reliably than prompt wording does.
How to switch effort in the Claude apps
In claude.ai and the Claude apps, click the model name next to the send button, choose Effort, then pick Low, Medium, High, Extra high or Max, according to Anthropic's help center. Each model's recommended level is labeled "Default" in the menu. You can change effort in the middle of a conversation, and the change applies from Claude's next reply.
The help center also warns that higher effort uses your plan's allowance faster. High and above take longer and use more tokens, while Low and Medium "stretch your usage further." On a Pro plan, that is a reason to keep medium as your default and use high only for specific tasks.
How to set effort in Claude Code
Claude Code also starts Opus 5.5 at medium, and it needs version 2.1.280 or later. The model configuration docs list several ways to change it:
/effortopens a slider./effort highsets the level directly, and/effort autoclears your saved level so the default comes back.- The
/modelpicker has an effort slider. Use the left and right arrow keys. claude --effort highsets the level for one session.- The
CLAUDE_CODE_EFFORT_LEVELenvironment variable overrides everything else. - A per-model level saved under
modelSettingsin your user settings. An older top-leveleffortLevelkey doesn't apply to Opus 5.5, so a saved setting from Opus 5 won't carry over.
Typing ultrathink in a prompt asks for deeper reasoning on that one turn without changing your session setting. Claude Code treats "think hard" and "think more" as ordinary text, not commands.
How to set effort in the API
In the Messages API, set output_config.effort:
{
"model": "claude-opus-5-5",
"max_tokens": 64000,
"output_config": { "effort": "high" },
"messages": [{ "role": "user", "content": "..." }]
}
If you leave the field out, the request runs at medium. Sending "medium" gives exactly the same behavior. Earlier Opus models defaulted to high, so code that relied on the default now runs one level lower unless you set effort yourself. Thinking tokens count toward max_tokens, so leave room for them. Anthropic says 128,000 tokens, the model's maximum, has worked well for long agentic coding turns.
Changing the top-level effort between requests invalidates the prompt cache for the conversation. To change effort partway through a conversation and keep the cache, use per-message effort, which is in beta.
Compared with other Claude models
If you're choosing between models as well as levels, see our Claude Opus 5.5 vs Sonnet 5.5 comparison and our Claude Fable 5.2 vs Opus 5.5 comparison. Anthropic's cost page notes that on its SWE-bench Pro subset, Opus 5.5 at low solved about as many tasks as Sonnet 5.5 at its default high (87.4% vs 87.6%), for about 7% more per solved task. Its new guide also says that pushing Sonnet 5.5 to Extra high or Max brings its cost per task close to Opus 5.5.
Frequently asked questions
What is the default effort for Claude Opus 5.5?
Medium. A Messages API request that leaves out effort runs at medium, and the Claude apps and Claude Code also start Opus 5.5 at medium. Earlier Opus models, including Opus 5, defaulted to high.
What is the best effort level for Claude Opus 5.5?
Anthropic recommends starting at medium and raising effort only where your own tests show a gain. High suits bug fixes and work where edge cases matter, and low suits quick exchanges and subagents.
How do I change effort for Claude Opus 5.5 in Claude Code?
Run /effort to open a slider, or /effort followed by a level such as high. You can also launch with claude --effort high, set CLAUDE_CODE_EFFORT_LEVEL, or use the effort slider in the /model picker.
Is Claude Opus 5.5 max effort worth it?
Usually not. Artificial Analysis measured about 25 minutes per task at max against 9.5 minutes at high, for a 7-point higher index score, and Anthropic says to reserve xhigh and max for work where you have measured a gain.
Sources
- EffortAnthropic
- Using Claude Opus 5.5 as your daily driver for knowledge workAnthropic
- Prompting Claude Opus 5.5Anthropic
- Optimizing for cost and intelligenceAnthropic
- What's new in Claude Opus 5.5Anthropic
- Model configuration (effort levels)Anthropic
- Change the model, effort and thinking settingsAnthropic
- Claude Opus 5.5: intelligence, performance and price analysisArtificial Analysis
- Anthropic says outdated prompting habits are slowing down Claude Opus 5.5The Decoder

