What is reasoning effort?

A setting on reasoning models (low, medium, high, and on some models minimal, xhigh or max) that controls how much the model thinks before answering. Higher settings spend more output tokens on reasoning, which usually raises scores and always raises cost and wait time.

Does high reasoning effort make a model better?

Usually, but not always. Of the 19 models measured at more than one setting as of 21 Oct 2026, the dearest setting outscores the cheapest for 18; the median gain is 5.8 index points.

How much more does high reasoning effort cost?

For the same batch of work, the dearest setting costs a median 5.5× the cheapest across the 19 models measured here, as of 21 Oct 2026. The per-token price does not change; the model just writes many more tokens.

Which reasoning effort should I use?

Start at the lowest setting that clears your quality bar and only step up where it measurably helps. Models with a high points-per-doubling figure on this page reward turning the dial; models with a low one mostly bill you for thinking that does not show up in the score.

Reasoning effort compared: what more thinking costs

As of 21 Oct 2026, across the 19 models measured at more than one reasoning effort, the highest setting costs a median 5.5× the lowest and adds a median 5.8 index points; DeepSeek V4.1 Flash gains the most per doubling of spend (11.7 points) and Qwen3.8 27B the least (-5.4).

Reasoning effort (also sold as a thinking budget or reasoning level: low, medium, high and so on) sets how much a model reasons before it answers. More reasoning means more output tokens, so it changes both the score and the bill. BenchLeader ranks each setting as its own configuration, and this page sets them side by side.

Most frontier models let you choose how hard to think, and the setting is usually presented as a quality dial. It is also a price dial, and the two do not move together. Every figure here is what it cost Artificial Analysis to run its whole Intelligence Index at that setting — a fixed batch of work, billed at list price — so a model that reasons at length pays for it in a way its per-token price never shows.

The spread is wide. DeepSeek V4.1 Flash buys 11.7 index points for every doubling of spend; Qwen3.8 27B buys -5.4. Turning Qwen3.8 27B all the way up costs 2× its cheapest setting and moves it backwards by 7.1 points.

19 models whose settings span at least a doubling in cost · data as of 21 Oct 2026

DeepSeek V4.1 Flash

11.7 points per doubling · 2.3× for +14.0

GLM 5.3

6.5 points per doubling · 2.4× for +8.0

GPT-5.6 Terra

4.5 points per doubling · 10× for +14.8

Claude Fable 5.1

3.3 points per doubling · 3.2× for +5.5

Claude Opus 5

3.1 points per doubling · 5.3× for +7.5

Claude Sonnet 5.5

2.8 points per doubling · 16× for +11.3

Claude Opus 5.5

2.3 points per doubling · 11× for +8.0

GPT-6 Sol

2.3 points per doubling · 7.9× for +6.9

GPT-6 Astra

2.3 points per doubling · 4.0× for +4.6

GPT-6 Luna

2.3 points per doubling · 15× for +8.9

GPT-5.6 Luna

2.2 points per doubling · 18× for +9.0

GPT-6.1 Sol

2.0 points per doubling · 5.5× for +5.0

GPT-5.5

2.0 points per doubling · 2.9× for +3.1

GPT-5.6 Sol

1.9 points per doubling · 7.6× for +5.7

Claude Haiku 5.5

1.9 points per doubling · 8.7× for +5.8

Claude Sonnet 5

1.6 points per doubling · 10× for +5.3

Grok 4.6

1.4 points per doubling · 4.7× for +3.2

Grok 4.7

0.4 points per doubling · 3.0× for +0.7

Qwen3.8 27B

-5.4 points per doubling · 2.5× for -7.1

How to read this

Points per doubling is index points gained each time the bill doubles. Spending scales geometrically here — each setting tends to cost a multiple of the last rather than a fixed amount more — so a per-doubling figure compares models whose price ranges are nothing alike. A high number means the dial is worth turning; a low one means you are paying for thinking that does not show up in the score.

Only models Artificial Analysis has run at more than one setting appear, because the cost figure comes from those runs. It is one publisher’s harness on one benchmark suite: a model that reasons usefully on work unlike theirs will look worse here than it is.

Questions people ask

What is reasoning effort?: A setting on reasoning models (low, medium, high, and on some models minimal, xhigh or max) that controls how much the model thinks before answering. Higher settings spend more output tokens on reasoning, which usually raises scores and always raises cost and wait time.
Does high reasoning effort make a model better?: Usually, but not always. Of the 19 models measured at more than one setting as of 21 Oct 2026, the dearest setting outscores the cheapest for 18; the median gain is 5.8 index points.
How much more does high reasoning effort cost?: For the same batch of work, the dearest setting costs a median 5.5× the cheapest across the 19 models measured here, as of 21 Oct 2026. The per-token price does not change; the model just writes many more tokens.
Which reasoning effort should I use?: Start at the lowest setting that clears your quality bar and only step up where it measurably helps. Models with a high points-per-doubling figure on this page reward turning the dial; models with a low one mostly bill you for thinking that does not show up in the score.

Cite as: BenchLeader, “Reasoning effort compared”, https://www.benchleader.com/effort, data as of 21 Oct 2026.