Published: 2026-09-24
Analysis & perspective

Opus 5.5 at every effort level on one /goal build: $3.91 at low, $50 at max — and xhigh won

Chapters / key moments (click to jump — plays here on the page)

The clearest effort-level data we've seen for Opus 5.5. One long autonomous task (turn 105 GB of event recordings into an explorable 3D conference) was run through /goal at every level, with every run measured. Opus 5.5 defaults to medium in Claude Code (model config docs).

Source video

"I Had Opus 5.5 Build me the Same App at Every Effort Level" by Nate Herk — Watch on YouTube →

The numbers

Effort	Runtime	API-equivalent cost	Tokens	Checks	Questions asked
low	16 m 43 s	$3.91	191K	22	0
medium	1 h 13 m	$12.44	419K	23	0
high	1 h 07 m	$16.31	~559K	22	1
xhigh ("extra")	~1 h 30 m	$25.92	~732K	34	0
max	2 h 28 m	$50.38	1.18M (auto-compacted)	51	0
ultracode	1 h 35 m	$18.69	as reported	42	0

Numbers as read out in the video. He ran on a subscription and computed API-equivalent cost. Some token figures are garbled in the auto-captions and are shown approximately. No run used subagents.

What the quality looked like

low: working but off-brand, still images instead of video, glitching NPCs.
medium: used his brand guidelines, played real session video, NPCs reacted. "For a lot of my knowledge work, medium works just fine."
high: best loading screen, working VIP flow and afterparty, closed captions on stage.
xhigh: his winner. You could talk to NPCs, sit in any session, and it had the best physics. It cost about half of max in time and money.
max: slowest and most expensive, with worse walking and more bugs. "Way too much for not enough good."

Ultracode didn't behave like ultracode

Per the Claude Code workflows docs, /effort ultracode makes Claude plan a workflow for every substantive task in the session. In his runs it never started one. It behaved like xhigh with more checks. He doesn't know whether that's a harness bug or Opus 5.5 declining workflows, so if you rely on ultracode, check that workflows actually start. See our dynamic workflows FAQ.

How to use this

/effort xhigh        # low | medium | high | xhigh | max | ultracode | auto
/effort status
/goal <condition>     # keep working across turns until the condition is met

Start at the default (medium) and step up only when the task needs deep reasoning over a lot of material, as Anthropic's own prompting guide suggests.
For long autonomous /goal runs, xhigh was the sweet spot here. Max roughly doubled cost and time for a worse result.
Per the docs, max and ultracode are session-only. Everything else persists.

Background: what each effort level does. (Includes a sponsor segment.)

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Results from the video (speaker's figures, subscription run priced as API)

low:    16m43s, $3.91, 191K tokens, 22 checks, 0 questions
medium: 1h13m, roughly $12, 419K tokens, 23 checks, 0 questions
high:   1h07m, $16.31, roughly 560K tokens, 22 checks, 1 question
extra:  1h30m, $25.92, 34 checks, 0 questions
max:    2h28m, $50.38, 1.18M tokens (hit an auto-compact), 51 checks, 0 questions

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Give every effort level the identical /goal
The prompt asked for a walkable third-person 3D world of a tech conference built from a folder of 105 GB of event recordings, judged on creativity, design, physics and feel. The same prompt ran at low, medium, high, extra, max and the top level (ultra).
Start at medium and move in either direction
The speaker relays Anthropic's guidance to start on medium effort and scale up or down as needed. In his run medium was already a good output.
Compare more than the output
He tracked run time, API-equivalent cost, total tokens, number of verification checks and number of questions the agent asked.

Gotchas

Cost and time did not rise in a straight line: high was slightly faster than medium. Token use and cost did climb at every step.
Quality did not climb in a straight line either. The speaker rated low weakest (unbranded, people vanishing, still images instead of video), medium and high good, extra his favourite, and max worse than extra: bad walking animation, glitches, and seats that could not be used despite 51 checks.
Only the high run asked a clarifying question; the others asked none.
This is one prompt, one run per level and one person's judgement, so treat it as an illustration rather than a benchmark.
The video includes a sponsor segment for Hostinger's editor connector; it is not part of the comparison.