`reasoning_effort` sets how much hidden thinking the model does before it answers, and that thinking is billed as reasoning tokens at the output rate [1]. So the rung you pick is a spend dial, not a mode switch, for every value except one.
The one exception is `none`, the only value that zeroes reasoning tokens [5]. It also fails 17 of the 33 verified runs [4]. Every other rung scored 33 of 33 [6]. The accuracy gap sits in that single step rather than spreading across the ladder, and that shape is what makes the pricing argument work: if `low` and `max` return the same answer on all 11 tasks [2], the extra tokens at the top are pure cost [3].
The 11 task answers were computed by exhaustive search on the author's own machine first, so the answer key cannot be wrong [7]. The tasks were things like a rule applied 40 times in a row, a counting problem under three constraints, a knapsack, a base conversion, and 7 to the power 222 modulo 1000 [8]. Those are deterministic, verifiable, short-horizon problems. For the ladder-buys-nothing conclusion to transfer to your workload, your tasks would have to be similar: single-shot, checkable, and already inside the model's competence at `low`. Long agent loops with tool calls are exactly the case this sweep does not cover, and they are the case OpenAI leads with, claiming Astra is ahead on agent-style work [11].
Then there is `disabled`, which is a rung the docs never mention and which the API accepts on Astra alone [9]. It spends 243 reasoning tokens against `low`'s 151 and costs 54% more per correct answer at the same accuracy [10]. Naming a setting `disabled` when it disables nothing is the kind of thing you only find by sending every value and reading the response codes.
The docs and the API disagree in both directions. The request-shape check advertises seven values including `none` and `minimal` [12], while the model page lists five and states the model does not support `none` [13]. `minimal` is advertised and rejected on every model [14]. `none` works despite the docs saying it does not [14].
The line that connects the benchmarks to the bill is OpenAI's own: "Evaluation scores are the maximum at any effort" [15]. The Artificial Analysis leaderboard lists each model once per effort setting, and Astra scores 54 at `xhigh` against 55 at `max` [16]. One point of aggregate score sits on top of a rung that, on these 11 tasks, changed nothing. On the same leaderboard's v4.2 version, Claude Fable 5.1 is first at 57 and Astra third at 55 [17], and OpenAI's own table puts Astra behind Fable 5.1, Opus 5 and Fable 5 on that row [18]. The Humanity's Last Exam row, where Astra trails Fable 5.1 by 7.8 points, appears in the table and not in the prose [19].
One more caveat on the security numbers: OpenAI says the cybersecurity scores were produced "without production safeguards", and the shipping model "will refuse" proof-of-concept exploit tasks [20]. That column reflects a configuration no buyer can actually run.
Cross-model, the same list prices put Astra at 1.57x GPT-5.6 Sol per correct answer at `low`, and 2.57x at `max`, on work both models get right [21]. Divide those and the effort ladder accounts for roughly 1.64x of the gap by itself [22]. The default worth defending is `low`, with `none` reserved for work where a wrong answer is cheap.
