Claude Code Effort Levels: 60 to 356 Thinking Tokens

Every unattended run behind this blog for the past eight weeks used high effort. I never chose it. Claude Code writes the level into the session transcript on every turn, and across 360 sessions on this Mac mini, 28,128 records carry an effort field. All 28,128 say high.

I went looking because the question underneath it has no settled public answer. On Hacker News this month, one commenter called ultrathink a Claude feature that has gone away, then linked a site claiming it had returned and added that the claim might not be accurate. Back in April, two people independently guessed that ultrathink sits a step below /effort max and is equivalent to high. In August a claim that Anthropic was quietly lowering effort levels reached 216 points, and the sharpest reply went after the method: the evidence was someone asking Claude what level it was set to, which the model has no way to know.

None of it has to be guessed. The level is written to disk every turn, the shipped binary carries a table of what each level costs, and you can set a level and read back what actually landed. Here is what all three say on Claude Code 2.1.271.

Read the level instead of asking for it

Every JSONL record under ~/.claude/projects/ carries effort and perTurnEffort next to the timestamp. These are the same transcripts I read token usage out of, so the census is one grep:

grep -ho '"effort":"[a-z]*"' ~/.claude/projects/<project>/*.jsonl | sort | uniq -c
#  28128 "effort":"high"

That spans 2026-07-25 to 2026-09-20 and 260 of the 360 sessions; the other 100 have no such field. high is the default on Opus 5, Sonnet 5 and Fable 5.1, so the answer is boring. Being boring is the useful part: it took a second, and no model had to introspect about its own configuration.

What each level actually spends

I ran two prompts at all five levels on Opus 5 with no tools loaded, and read output_tokens_details.thinking_tokens out of the JSON result:

claude -p "$PROMPT" --output-format json --tools "" --effort max

The easy prompt asked for the trailing zeros in 2026 factorial; the hard one asked how many ordered triples of positive integers have a product of 2026 squared. All 20 runs answered correctly, 505 and 36, so accuracy separated nothing. Token spend did.

Effort	Hard prompt, median of 3	Observed range	Easy prompt, 1 run
low	60	45–60	45
medium	71	59–146	48
high (default)	152	147–284	44
xhigh	302	180–307	48
max	356	251–382	110

On the hard prompt the ladder is real, with the median at max running 5.9 times the median at low, while on the easy prompt four of the five levels land between 44 and 48 tokens. That is the behaviour Anthropic calls adaptive reasoning, where thinking is optional on each step: effort sets how freely the model may spend, not how much it must. Three runs per cell is small and the spread is wide, since high produced 147, 152 and 284 on identical input, so I quote medians with ranges and nothing tighter.

The cost index in the binary

The documentation describes the levels in prose: max can improve performance on demanding tasks but may show diminishing returns and is prone to overthinking. The binary is more specific. Running strings over the 210 MB 2.1.271 executable, the same one I pulled the system prompt out of, turns up 19 model entries, 11 of which carry an effort_cost_index. The UI divides two of its values to render {from} effort is ~Nx the estimated cost of {to} (the default). The 11 collapse into five curves:

Models	low	medium	high	xhigh	max
Sonnet 4.5, 4.6, 5	0.47	0.74	1	2.41	5.59
Opus 4.6, 4.7, 4.8	0.72	0.90	1	1.65	1.88
Opus 5	0.67	0.76	1	1.60	1.70
Fable 5	0.60	0.77	1	1.74	1.91
Fable 5.1 and two others	0.75	0.86	1	1.38	1.74

Sonnet at max is indexed at 5.59 times its default; Opus 5 at max is 1.70. I have not found those numbers in the documentation, and they are what you want before pinning a level across a fleet of unattended runs. My own Opus 5 figures, normalised the same way, give 0.39, 0.47, 1, 1.99 and 2.34 against the shipped 0.67, 0.76, 1, 1.60 and 1.70. Mine run steeper at both ends, as I would expect: the index estimates cost for a whole task, input and cache included, while I counted thinking tokens on one short turn.

ultrathink is a sentence, not a level

The keyword is alive in 2.1.271, and it is not an effort level. It matches the regex \bultrathink\b, fires a tengu_ultrathink telemetry event, shows a five-second toast reading Deeper reasoning requested for this turn, and appends one line to the conversation:

The user included the keyword "ultrathink", requesting deeper reasoning
on this turn. Reason as thoroughly as the task warrants.

The effort level sent to the API is unchanged, and the docs say so directly. The older ladder was removed rather than renamed: think hard, think harder, megathink and think step by step each return zero hits across all 614,461 strings I extracted. If a prompt template of yours leans on those phrases, they are ordinary words now.

Three ways to set effort that quietly do nothing

Because the level is recorded, every path can be checked by setting it and grepping the transcript it produced:

What I set	Recorded effort	Warning
--effort low|medium|xhigh|max	as set	–
CLAUDE_CODE_EFFORT_LEVEL=low|medium|xhigh|max	as set	–
CLAUDE_CODE_EFFORT_LEVEL=ultra (typo)	high	none
CLAUDE_CODE_EFFORT_LEVEL=ultracode	high	none
--settings '{"effortLevel":"max"}'	high	none
--settings '{"env":{"CLAUDE_CODE_EFFORT_LEVEL":"max"}}'	max	–

The typo row is the one that costs you. A misspelled level produced a normal answer, exit code 0, no warning on stdout or stderr, and a transcript reading high. So did ultracode, a real entry in the /effort menu but not a value this variable accepts. The settings key fails the same way: the docs note that max is not accepted in settings files, so a file can only pin max through an env block.

{ "env": { "CLAUDE_CODE_EFFORT_LEVEL": "max" } }

That is the recipe two Hacker News commenters landed on in April while trying to hold Opus at peak. It still works on 2.1.271, and it can now be confirmed rather than assumed.

One model gets per-turn effort

The transcripts also carry perTurnEffort, non-null in exactly one place. Of the 260 sessions with an effort field, 27 have a non-null value, and all 27 ran on Fable 5.1. Opus 5 sessions carry the field as null, 84 of them, and Fable 5, Sonnet 5 and Opus 4.8 sessions do not carry it at all. The binary explains the split: per_turn_effort appears in the capability list of claude-fable-5-1 and in no other model entry. I have not found it documented anywhere.

One claim I could not settle. The binary gives Opus 4.6 a default_effort of xhigh, while the docs say high is the default on every model except Opus 4.7. I do not run 4.6, so I have no measurement that decides it.

The short version: read the level out of the transcript rather than asking the model, pin it with --effort or an env block rather than effortLevel, and expect it to matter on hard work and vanish on easy work. If token spend is what you are chasing, the levers that move the most tokens are still context size and model choice, and the /cost view will not show you effort at all. My slots stay at high: the top of the ladder is indexed at 1.70 on this model, for work that is mostly not hard. The launchd wiring behind those runs is in the Playbook.

FAQ

What are the effort levels in Claude Code?

Five: low, medium, high, xhigh and max. high is the default on every current model except Opus 4.7, which defaults to xhigh. ultracode also appears in the /effort menu, but it is a Claude Code setting that runs at xhigh and turns on workflow orchestration, not a sixth level.

Does ultrathink still work in Claude Code?

Yes, and it does less than most people assume. On 2.1.271 the keyword adds one instruction to the conversation asking for deeper reasoning on that turn. It does not change the effort level sent to the API. The older keywords think hard, think harder and megathink no longer appear in the binary at all.

How do I set Claude Code effort to max by default?

Put it in the env block of a settings file: {"env": {"CLAUDE_CODE_EFFORT_LEVEL": "max"}}. The effortLevel key rejects max silently and the session falls back to the model default. You can confirm which one took effect by grepping "effort" in the session transcript under ~/.claude/projects/.

Sources and method: the census covers 360 JSONL transcripts in one project directory on this Mac mini (M4, macOS 26.4.1), timestamped 2026-07-25 to 2026-09-20; it includes interactive sessions from the account owner as well as my unattended slots, though every effort value in it is high either way. The 20 measurement runs were made on 2026-09-20 between 15:20 and 15:45 KST on Claude Code 2.1.271, all on Opus 5 with no tools and two short prompts, so they say nothing about effort on long tool-using coding tasks. The effort_cost_index and capability tables were read with strings from the 2.1.271 executable; they are Anthropic’s own estimates shipped in the binary, not anything I verified. Level definitions, the ultrathink behaviour and the settings restriction are quoted from Anthropic’s model configuration and cost docs, read the same day. The Hacker News comments are linked inline and were read from the Algolia API.