Project benchmark

LiveBench 2.0 task results & run costs

LiveBench · 31 configurations with published scores. Review each metric and its test conditions separately.

Report identity & conditions

Data as of2026-10-27· Report published—. See individual records for run dates and test conditions.

Leaderboard updated2026-10-25

Each run retains a 0–100 score for each of seven task categories. The original report assigns zero to execution failures that remain unresolved after retries. Missing cells remain missing; we do not fill them with zero.
The overall score follows the current official methodology and version configuration: long tasks ×2.0, reasoning ×1, coding ×1.5, browsing ×1, data analysis ×1, research ×1, documents/design ×1.5. Only configurations with all seven scores enter the comparison; versions are kept separate.
Costs cover all calls to the tested model at provider list prices, including cached input rates. Fixed initial state and supporting framework calls are excluded. Each task retains its total cost and count: USD/task = total cost ÷ task count. This is neither cost per successful task nor a current API unit price.
Overall cost per task is the sum of costs divided by the sum of task counts across all seven categories. It is provided only when every category has complete costs and samples, avoiding direct comparisons between different task mixes.
Models are shown under the report’s configuration identifiers. Shared names or mutable API aliases do not prove an exact snapshot. Test dates, parameters and settings follow each run’s documentation.
Results come from the official LiveBench leaderboard. Follow the report link for methodology and original versions. We do not reproduce test questions, answers, benchmark code or API prices without a defined scope.
The release notes retain an older equal-weight statement that conflicts with the current methodology and version configuration. We use the explicit weighted formula in the latter two sources and retain this source discrepancy.
Benchmark version: 2026-10-25; leaderboard updated: 2026-10-25. Neither is an individual test date or an official model release date.

Left metricRight metric

Compare metrics within the same report. Each chart explains its direction and scope.
See the details for full values and run conditions.

Overall score

Weights for this version: long tasks ×2.0, reasoning ×1, coding ×1.5, browsing ×1, data analysis ×1, research ×1, documents/design ×1.5. Overall score = sum of weighted scores ÷ sum of weights, and requires all seven scores.: showing 8 of 31 records with values. See details for configurations and dates.

Chart data · 8 entries

Name	points
claude-fable-5.1	85.69
gpt-6-astra	85.64
claude-opus-5.5	85.14
gpt-6.1-sol	84.07
claude-sonnet-5.5	80.04
gemini-4-argon	79.3
gpt-6.1-terra	77.79
muse-wonder-1.0	74.15

Long tasks score

The run’s mean score for this task, from 0 to 100. It is not ranked directly against other benchmarks or tasks.: showing 8 of 31 records with values. See details for configurations and dates.

Chart data · 8 entries

Name	points
claude-fable-5.1	77.9
gpt-6-astra	77.5
claude-opus-5.5	72.8
gpt-6.1-sol	68.7
claude-sonnet-5.5	63.4
gemini-4-argon	61.2
muse-wonder-1.0	52.6
gpt-6.1-terra	48.6

Full results & model versions

31 / 31configurations have published scores in this scope. Scored configurations are listed; missing values remain “Not provided.” Different run configurations of the same model are listed separately.