AI Agent StoreLow
Tested setting-6.7 pts vs Extra high
$0.603 / source task
96 s median duration
Anthropic · REASONING EFFORT GUIDEClaude Opus 4.8
Claude Opus 4.8
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 Claude Opus 4.8
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 TESTExtra high
Lowest measured cost within 1.0 points of the best tested analytics score.
18.7 / 100 · $1.2 / 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$0.603
Medium$0.759
High$0.8574
Extra highNear-best pick$1.2
Max$1.78
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-3.7 pts
Noticeable gap · B scored lower
Measured cost · B / A1.49×
Same benchmark workload
Median duration · B / A1.73×
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.
When more thinking scored lower
Medium → High-3.0 pts17.0 → 14.0 on analytics score
Extra high → Max-3.7 pts18.7 → 15.0 on analytics score
EVERY TESTED MODE
5 modes · untested settings are left blankResults for each tested setting
12.0 / 100
Medium
Tested setting-1.7 pts vs Extra high
$0.759 / source task
133 s median duration
High
A better tradeoff is available-4.7 pts vs Extra high
$0.8574 / source task
187 s median duration
Extra high
Suggested setting0.0 pts vs Extra high
$1.2 / source task
251 s median duration
Max
A better tradeoff is available-3.7 pts vs Extra high
$1.78 / source task
434 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 | 12.0 | $0.603 | 96.2 | — | Opus 4.8 · Low |
| Medium | 17.0 | $0.759 | 133.4 | — | Opus 4.8 · Medium |
| High | 14.0 | $0.8574 | 186.7 | — | Opus 4.8 · High |
| Extra high | 18.7 | $1.2 | 251.0 | — | Opus 4.8 · XHigh |
| Max | 15.0 | $1.78 | 433.9 | — | Opus 4.8 · Max |
SCALE THE OBSERVED COST
Effort cost difference calculator
What would A versus B cost if you repeated this benchmark’s average workload?
A · Extra high$1,200.64
B · Max$1,783.55
Difference$582.92more 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 Claude Opus 4.8. Rates start from the current model snapshot.
A · Extra high$100estimated monthly cost
B · Max$200estimated monthly cost
Mode B costs $100 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.
Claude Opus 4.8 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
- 12.0 · $0.603 / task
- Medium
- 17.0 · $0.759 / task
- High
- 14.0 · $0.8574 / task
- Extra high · pick
- 18.7 · $1.2 / task
- Max
- 15.0 · $1.78 / task
Near-best pick: Extra high within 1 point of the best score. Highest-to-lowest score gap: 6.7 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 | 12.0 | $0.603 | 96.2 | Opus 4.8 · Low |
| Medium | 17.0 | $0.759 | 133.4 | Opus 4.8 · Medium |
| High | 14.0 | $0.8574 | 186.7 | Opus 4.8 · High |
| Extra high | 18.7 | $1.2 | 251.0 | Opus 4.8 · XHigh |
| Max | 15.0 | $1.78 | 433.9 | Opus 4.8 · Max |
General tasks · LiveBench 2026-01-08 · Thinking
AI Agent Store task-weighted category mix. An editorial selection aid, not a measured agent success rate; weights match our published Agent Fit methodology.
- Low
- 66.6
- Medium
- 71.0
- High
- 74.1
- Extra high · pick
- 75.5
Near-best pick: Extra high within 1 point of the best score. Highest-to-lowest score gap: 8.8 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 | Overall benchmark | General agent fit | Coding agent fit | Research agent fit | Writing & support fit | Reasoning | Coding | Agentic Coding | Mathematics | Data Analysis | Language | Instruction following | Exact source configuration |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Low | 70.2 | 66.6 | 58.6 | 75.6 | 70.1 | 83.5 | 76.5 | 38.3 | 80.5 | 76.4 | 77.3 | 59.1 | claude-opus-4-8-low-effort |
| Medium | 74.2 | 71.0 | 68.4 | 76.4 | 71.0 | 84.5 | 78.9 | 58.3 | 81.6 | 77.2 | 80.4 | 58.1 | claude-opus-4-8-medium-effort |
| High | 75.9 | 74.1 | 69.7 | 79.7 | 74.6 | 89.7 | 76.4 | 60.0 | 82.0 | 78.0 | 81.6 | 63.4 | claude-opus-4-8-high-effort |
| Extra high | 77.2 | 75.5 | 71.0 | 80.6 | 76.4 | 89.7 | 79.3 | 60.0 | 84.3 | 78.3 | 81.4 | 67.4 | claude-opus-4-8-xhigh-effort |
Math, code & science · OckBench
Correct answers out of 200 problems. Domain proportions are 100 math, 60 coding and 40 science.
- Low
- 73.5
- Medium
- 86.0
- High
- 86.0
- Extra high · pick
- 91.5
- Max
- 90.0
Near-best pick: Extra high within 1 point of the best score. Highest-to-lowest score gap: 18.0 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 | Overall accuracy | Mathematics | Coding tests | Science | Average output tokens | Exact source configuration |
|---|---|---|---|---|---|---|
| Low | 73.5 | 80.0 | 70.0 | 62.5 | 3,652 | opus-4.8-low |
| Medium | 86.0 | 88.0 | 86.7 | 80.0 | 7,712 | opus-4.8-medium |
| High | 86.0 | 91.0 | 81.7 | 80.0 | 8,870 | opus-4.8-high |
| Extra high | 91.5 | 90.0 | 93.3 | 92.5 | 16,876 | opus-4.8-xhigh |
| Max | 90.0 | 84.0 | 100.0 | 90.0 | 36,544 | opus-4.8-max |
Reasoning effort, explained
Which Claude Opus 4.8 thinking level is most efficient?
Extra high is the lowest-cost option within 1 point of the best analytics score in Hex DataBench v1.1. Max trails the best tested score by 3.7 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 Claude Opus 4.8 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 Claude Opus 4.8?
In Hex DataBench v1.1, Medium to High lowers the observed score by 3.0 points; Extra high to Max lowers the observed score by 3.7 points. The overall tested range is 6.7 points. These are observed differences, not proof that extra thinking caused the decline.
How much more does high effort cost for Claude Opus 4.8?
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.