AI Agent StoreLow
Tested setting0.0 pts vs Low
$0.0108 / source task
64 s median duration
OpenAI · REASONING EFFORT GUIDEGPT-6 Luna
GPT-6 Luna
How much thinking?
Compare Low vs Medium vs High vs Extra high vs Max where tested. Find the effort setting that fits your task, and see how much extra reasoning changes the result.
SAME MODEL. DIFFERENT THINKING.
Choose a thinking level for GPT-6 Luna
Hex DataBench v1.1
Published blended DataBench score, on a 0–100 scale. Judge-based analytical task performance; not a general intelligence or guaranteed success rate.
Analytics in the Hex agent harness · Leaderboard updated: 2026-09-22 · Reviewed 2026-10-15Check the source ↗SUGGESTED SETTING TO TESTMax
Lowest measured cost within 1.0 points of the best tested analytics score.
52.3 / 100 · $0.058 / source taskExact best1.0 pts10 pts
Set 0 for the highest score. Increase the gap to explore cheaper or lighter options.
THE EFFORT CURVE
Does more thinking help?
● Near-best pickOrange = score decreased
Scores out of 100 · Full 0–100 scale · tap a point to select mode B. Cyan band: within 1.0 points of the best tested score.
WHAT YOU SPEND
Cost of more effort
Measured average USD per benchmark task.
Low$0.0108
Medium$0.0191
High$0.0258
Extra high$0.0396
MaxNear-best pick$0.058
Measured in this study with its particular tools and setup. Your task mix, tools, caching and endpoint can change the result.
Score difference · B minus A+22.0 pts
Large gap · B scored higher
Measured cost · B / A5.37×
Same benchmark workload
Median duration · B / A5.56×
Measured with the study’s testing setup
We describe gaps under 1 point as small, 1–under 5 as noticeable, and 5 or more as large. These are reading aids, not proof that a difference would hold up in repeated tests. Equal scores can still hide different mistakes.
EVERY TESTED MODE
5 modes · untested settings are left blankResults for each tested setting
30.3 / 100
Medium
Tested setting+12.3 pts vs Low
$0.0191 / source task
117 s median duration
High
Tested setting+15.7 pts vs Low
$0.0258 / source task
175 s median duration
Extra high
A better tradeoff is available+15.7 pts vs Low
$0.0396 / source task
189 s median duration
Max
Suggested setting+22.0 pts vs Low
$0.058 / source task
353 s median duration
Full benchmark table & exact configuration IDs
Swipe inside the table for all measures. Figures from different study selections are not directly comparable.
| Effort | Analytics score | Cost / task | Median seconds | Output tokens | Exact source ID |
|---|---|---|---|---|---|
| Low | 30.3 | $0.0108 | 63.5 | — | GPT-6 Luna · Low |
| Medium | 42.7 | $0.0191 | 116.6 | — | GPT-6 Luna · Medium |
| High | 46.0 | $0.0258 | 175.4 | — | GPT-6 Luna · High |
| Extra high | 46.0 | $0.0396 | 188.7 | — | GPT-6 Luna · XHigh |
| Max | 52.3 | $0.058 | 353.2 | — | GPT-6 Luna · Max |
SCALE THE OBSERVED COST
Effort cost difference calculator
What would A versus B cost if you repeated this benchmark’s average workload?
A · Low$10.8
B · Max$57.96
Difference$47.17more with B
A source-workload scenario using average measured USD costs. Source prices, caching and tools are already reflected in those observed costs. Your agent’s costs may differ.
YOUR WORKLOAD
Reasoning-token cost calculator
Estimate the extra bill from thinking tokens. The starting token counts are editable examples, not measurements of GPT-6 Luna. Rates start from the current model snapshot.
A · Low$2estimated monthly cost
B · Max$4estimated monthly cost
Mode B costs $2 more per month with these assumptions.
This estimate uses the token counts and prices you enter. Excludes retries, tools, caching, long-context premiums, hosting and taxes. Includes thinking in output billing; do not also include it in “visible output.” Check endpoint limits separately. Changing mode does not automatically predict tokens or quality.
GPT-6 Luna effort benchmarks at a glance
Each card compares settings tested together in one study. Scores from different benchmarks or question sets measure different work and are not averaged together.
Analytics · DataBench v1.1
Published blended DataBench score, on a 0–100 scale. Judge-based analytical task performance; not a general intelligence or guaranteed success rate.
- Low
- 30.3 · $0.0108 / task
- Medium
- 42.7 · $0.0191 / task
- High
- 46.0 · $0.0258 / task
- Extra high
- 46.0 · $0.0396 / task
- Max · pick
- 52.3 · $0.058 / task
Near-best pick: Max within 1 point of the best score. Highest-to-lowest score gap: 22.0 points.
Explore this study ↗All results and the exact settings tested
A dash means the study did not report that result. This table compares only the versions and settings tested on the same questions in this study.
| Effort / budget | Analytics score | Measured USD / task | Median seconds / task | Exact source configuration |
|---|---|---|---|---|
| Low | 30.3 | $0.0108 | 63.5 | GPT-6 Luna · Low |
| Medium | 42.7 | $0.0191 | 116.6 | GPT-6 Luna · Medium |
| High | 46.0 | $0.0258 | 175.4 | GPT-6 Luna · High |
| Extra high | 46.0 | $0.0396 | 188.7 | GPT-6 Luna · XHigh |
| Max | 52.3 | $0.058 | 353.2 | GPT-6 Luna · Max |
Math and logic · 44-problem Codex study
Correct answers out of 88 observed runs on 44 fixed problems, repeated twice. Repeats are not 88 independent questions; reported confidence intervals overlap for the near-ceiling models.
- Medium
- 81.8
- High · pick
- 86.4
Near-best pick: High within 1 point of the best score. Highest-to-lowest score gap: 4.5 points.
Explore this study ↗All results and the exact settings tested
A dash means the study did not report that result. This table compares only the versions and settings tested on the same questions in this study.
| Effort / budget | Answers correct | Median seconds / task | Reported 95% accuracy interval | Exact source configuration |
|---|---|---|---|---|
| Medium | 81.8 | 9.1 | 72.5–88.5% | gpt-6-luna / medium / 44-problem Codex test |
| High | 86.4 | 9.0 | 77.7–92.0% | gpt-6-luna / high / 44-problem Codex test |
Download this study’s data (CSV) ↓
Data: Joonlab · Codex subscription study. CC BY 4.0 ↗. AI Agent Store aggregates the published rows and adapts their labels. These are the same 44 problems at both settings. A small difference, often driven by one question, is a reason to test your tasks rather than a universal effort recommendation. Duration is the source median including network and queue time; no API cost is estimated.
Five efforts · one arithmetic prompt
Correct answers in three repeats of one prompt: count ordered integer pairs from 1 to 60 whose product is a perfect square; expected answer 158. One miss changes the score by 33.3 points. This is not a general reasoning benchmark.
- Low
- 66.7
- Medium
- 66.7
- High · pick
- 100.0
- Extra high
- 66.7
- Max
- 66.7
Near-best pick: High within 1 point of the best score. Highest-to-lowest score gap: 33.3 points.
Explore this study ↗All results and the exact settings tested
A dash means the study did not report that result. This table compares only the versions and settings tested on the same questions in this study.
| Effort / budget | Arithmetic runs correct | Median seconds / task | Average output tokens | Exact source configuration |
|---|---|---|---|---|
| Low | 66.7 | 19.9 | 863 | gpt-6-luna / low / reasoning prompt |
| Medium | 66.7 | 16.4 | 835 | gpt-6-luna / medium / reasoning prompt |
| High | 100.0 | 18.1 | 783 | gpt-6-luna / high / reasoning prompt |
| Extra high | 66.7 | 24.8 | 1,415 | gpt-6-luna / xhigh / reasoning prompt |
| Max | 66.7 | 41.0 | 1,841 | gpt-6-luna / max / reasoning prompt |
Download this study’s data (CSV) ↓
Data: Joonlab · Codex subscription study. CC BY 4.0 ↗. AI Agent Store aggregates the published rows and adapts their labels. One arithmetic question and three repeats per setting cannot establish general accuracy. The displayed output and reasoning tokens are means of raw runs; durations are medians. All five Sol6.1 settings answered this question correctly, so extra reasoning showed no accuracy gain on this example. These observations predate Ultrafast and do not compare service tiers.
Reasoning effort, explained
Which GPT-6 Luna thinking level is most efficient?
Max is the lowest-cost option within 1 point of the best analytics score in Hex DataBench v1.1. The tested score range is 22.0 points. This allows a score up to 1 point below the best result in the first listed study. Change the study, task and allowed score gap above to explore another comparison. It is a shortlist for testing your own workload.
Which GPT-6 Luna effort settings are compared?
Low, Medium, High, Extra high, Max appear in the collected studies. Each chart only includes modes tested together in that study. These published labels do not establish that every endpoint or application exposes every setting.
Does more effort improve GPT-6 Luna?
In the first listed Hex DataBench v1.1 sweep, the tested score range is 22.0 points and no adjacent effort increase lowers the score. This does not guarantee a gain on every task or with a different testing setup.
How much more does high effort cost for GPT-6 Luna?
Use the A/B selector and measured-cost calculator when the source reports dollar costs. For your own agent, enter average input, visible output and thinking tokens in the token calculator. Higher effort is adaptive, so the label itself cannot tell you the bill. Missing costs stay missing.