The bottom of the effort dial does not behave like a dial. At `low`, Simon Willison's pelican prompt produced 1,998 output tokens in 23.8 seconds for 10.017 cents, with no summarized reasoning in the transcript [9]. At `medium` it produced 1,977 tokens in 23 seconds for 9.912 cents, 21 tokens fewer than `low` [10]. Claude's output token count includes reasoning tokens [11], so Willison reads both runs as skipping reasoning entirely on this prompt [12]. `high` bought a short planning summary for 2,612 tokens, 29.6 seconds and 13.087 cents [13]. The steps that cost real money are `xhigh`, at 36,767 tokens, 7 minutes 51 seconds and $1.83 [14], and `max`, at 65,927 tokens, 13 minutes 54 seconds and $3.30 [15].
Divide cost by output tokens at each level and the rate is flat, around $50 per million at `low`, `high`, `xhigh` and `max` [1]. The prompt is one sentence, so input tokens are noise inside those totals. Divide tokens by wall clock and throughput is flat too, between 78 and 84 tokens per second [8]. Effort turns out to be a token budget on one meter, not a price tier and not a different serving path. `max` costs 33 times `low` because it writes 33 times as many tokens [2][3], and it takes 35 times as long [4]. In my context the useful position is `high`, because it is the last one before latency crosses from seconds into minutes [13][14].
The 52.6% is a much bigger jump than anything in those runs: 27.9 points over Fable 5 and 30.2 points over GPT-5.6 Sol on the same board [5]. It is also a score on a 0.1-versioned benchmark first announced on 27 August, five days before the model shipped [5][6], and Willison's reading of the rest of the announcement is that other benchmarks improved only slightly [6]. For that number to transfer, your tasks need to resemble the terminal-driven science tasks in that harness, the harness needs to score them the way you would score them, and you need to call the model at whatever effort level produced the score. The writeup does not name that level, and these runs show effort moving output volume by a factor of 33 on this same model [3].
Willison is explicit that he lost confidence in the pelican prompt as a proxy for general model quality back in July, and now uses it mainly for comparisons within a model family and for one prompt across reasoning levels [16]. That is the honest use of it, and it is the same discipline a five-day-old benchmark has not had time to earn. `max` gave him the best pelican he has seen from any Anthropic model [15], still with less flair than Gemini 3.7 Flash [17]. $3.30 for an SVG is a rounding error right up until it sits inside a retry loop.
