Complex terminal work spans coding, science, infrastructure, and other technical tasks, with executable checks.
94 model results
One row per model and compatible score measure. Mean = equal-weight arithmetic average of imported reports. Distinct configurations reported for the same model in a source table remain separate. Reports can overlap; the range is not a confidence interval. Unbundle to inspect each observation.
Published observations span 6.8 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#6 of 88 in this displayed comparison; 4.17 pp below the leader.
Published observations span 1.2 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#8 of 88 in this displayed comparison; 8.73 pp below the leader.
57.07% · Terminal-Bench 4.0 · AA harness · Publisher setting: High [1]
Published observations span 0.33 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#12 of 88 in this displayed comparison; 9.94 pp below the leader.
57.07% · Terminal-Bench 4.0 · AA harness · Publisher setting: Max [1]
Published observations span 0.03 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#13 of 88 in this displayed comparison; 10.09 pp below the leader.
Published observations span 5.9 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#17 of 88 in this displayed comparison; 12.47 pp below the leader.
Published observations span 0.8 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#19 of 88 in this displayed comparison; 14.94 pp below the leader.
Published observations span 0.02 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#20 of 88 in this displayed comparison; 15.16 pp below the leader.
Published observations span 4.02 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#25 of 88 in this displayed comparison; 27.26 pp below the leader.
Published observations span 2.6 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#29 of 88 in this displayed comparison; 29.01 pp below the leader.
33.84% · Terminal-Bench 4.0 · AA harness · Publisher setting: Max [1]
Published observations span 8.46 percentage points. Configuration and source differences contribute to this spread; it is not measured run-to-run uncertainty.
#30 of 88 in this displayed comparison; 29.1 pp below the leader.
Publisher-selected recent-model cohort from the public evaluation page. Model labels preserve effort and fallback configurations. Values are a dated snapshot, not a live feed. Consult the publisher’s methodology for the harness, task split and judging; do not merge this cohort with vendor-run results.
Values retain the model card’s original formatting and score units. HLE is without tools; the final three visual evaluations use tools. The dagger marks source qualifications: inspect the original report before treating those results as matched.
Reported values are percentages except GDPval-AA, which is Elo. Tool-enabled and partial scores keep their labels. Effort settings, fallback models, and source-run protocols vary; see the original footnotes.
A transcription of the October model card, including its external leaderboard snapshots. These are not uniformly matched runs. Effort, tools, hardware, frame budgets and task subsets differ. GraphWalks reports F1; OSWorld uses the offline subset and partial credit. Harvey refers to the archived Vals v1 protocol, not the remediated v1.1 or AA variant. Consult the linked methodology.
Launch-reported maxima across reasoning efforts, not a uniformly matched effort comparison. Research/API settings can differ from ChatGPT. Source footnotes disclose altered harnesses, model revisions and reproduced competitor results. A dagger marks cells with additional source qualifications. Preserve each row’s named subset and tools.
Source-reported model settings are max unless a cell says otherwise. The original page reports numbers without a uniform percent symbol; original formatting is retained. OSWorld includes two scoring rules in each cell: leaderboard ordering uses partial credit only. The internal composite and AA index use separate scales; their scores cannot be averaged.
Public source leaderboard captured 26 October 2026. Model names, configurations, native score units and source-reported uncertainty are retained. Version and task split follow this specific Vals methodology page. These are published observations, not evaluations run by RLTasks.
Terminal-Bench 4.0 · Kimi K3.1 Pro · complete release comparison · 2026-10-27
Published model comparison, 2026. Kimi K3.1 Pro · complete release comparison. Exact values from the publisher’s embedded full comparison table. Maximum effort, fallback settings and dated model versions remain in model labels. Blank cells stay missing. Source footnotes specify mixed harnesses and externally quoted comparisons; this is a launch-reported cohort rather than a new independent evaluation. Benchmarks · full embedded comparison table · row 1; source precision and any reported uncertainty retained. These are source-reported observations, not evaluations run by RLTasks.
TRACE THE SCORE · Per-model notes use published breakdowns where available.
Score vs. cost per test
Vals AI · 54 observations
Terminal-Bench 4.0 · Score · higher is better
Costs are Vals AI’s published Cost / Test in USD, paired with the score from the same result. They reflect that evaluation’s workload and configuration; they are not token prices or cost per successful task. Missing costs are omitted. Explore all reported results ↓
The dashed line traces the best observed score as reported cost increases. It describes this selection of reports, with their original configurations. It does not estimate unmeasured models or costs.
View plotted values and sources (54)
Terminal-Bench 4.0 · Terminal-Bench 4.0 · Score. Published measurements in their original score units.
These model reports include this benchmark. Their harnesses, task subsets, scoring policies, and resource budgets can differ; follow each report’s evaluation notes before joining the numbers.
Complex terminal work spans coding, science, infrastructure, and other technical tasks, with executable checks.
How to read the result
Treat this as a model-plus-agent measurement. Inspect failures that arise from planning, environment handling, and incorrect implementation separately.
Limits and comparison traps
The 4.0 task set and budgets differ from 2.1 and Hard. Results cannot be joined into a single trend without rerunning a common contract.
Epoch data and evaluation setting
Epoch AI publishes this benchmark in its capability collection. Imported rows retain the published score column and model configuration. Original authors and evaluation providers remain credited in the result references.
Epoch attribution and score coverage
Data is attributed to Epoch AI under CC BY 4.0, with original evaluation providers credited in the linked source snapshots. Scores are imported only where a published mapping identifies the field and its scale; missing or incompatible records are omitted rather than assigned zero.
Published evaluation
The linked publication documents the task and scoring procedure. Model results retain the source’s evaluation setting, native score units and available uncertainty.
Published evaluation
Exact values from the publisher’s embedded full comparison table. Maximum effort, fallback settings and dated model versions remain in model labels. Blank cells stay missing. Source footnotes specify mixed harnesses and externally quoted comparisons; this is a launch-reported cohort rather than a new independent evaluation.
Where this fits in an evaluation suite
The agent wrapper is part of the system. Pin tools, dependencies, timeouts, tests, and the repository revision. Compare contest coding separately from repository maintenance.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Exact values from the publisher’s embedded full comparison table. Maximum effort, fallback settings and dated model versions remain in model labels. Blank cells stay missing. Source footnotes specify mixed harnesses and externally quoted comparisons; this is a launch-reported cohort rather than a new independent evaluation.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.