Read model scores and agent benchmarks together.
A model result and an agent result describe different tested systems. These examples put their evidence next to the reading steps, while keeping model-only scores and complete model-plus-harness submissions in separate tables.
31 model results for the selected coding metric and 21 Terminal-Bench 4.0 submissions are available. Up to six examples from each are shown in separate tables; their scores are never merged.
Catalog snapshot · 12 rows shownRead the identity and setup before the score
Start with benchmark name, version and evaluation configuration. A model benchmark tests the setup its source describes. An agent benchmark also depends on the harness, its version, reasoning effort and allowed tools. The same model can appear in more than one submitted setup.
The agent examples use one benchmark version. Their pass rates are ordered only within that version. Model-only scores use a separate named coding metric and never share a ranking with agent pass rates. A common 0–100 display is not a common test.
Keep uncertainty and dates attached
A reported confidence interval and sample count give context to a score difference. Their absence means the source did not report them in the retained record. Do not treat a small gap as a measured advantage without inspecting the source’s uncertainty and test details.
Check dates show when evidence was collected. Evaluation dates say when the test ran, and release dates identify a different event again. A later check does not make an old evaluation new. Use the original result or submission to check the tested model identity and reproduce the relevant setup before applying the result to your task.
How this view is assembled
Selection: up to six model results for the selected LiveBench coding metric and up to six agent submissions for one retained benchmark version. One most recently checked model configuration is retained per model. Agent configurations remain distinct; unknown exact model identities receive no guessed profile link. No cross-metric or model/agent score is calculated.
Model benchmark examples
LiveBench Coding · 2.0 · 2026-10-25 (dataset 20261025); scores ordered only within this metric
| Model | Score | Benchmark | Tested setup | Sources and dates |
|---|---|---|---|---|
| Anthropic: Claude Fable 5.1 | 97.7%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
| Anthropic: Claude Opus 5.5 | 96.3%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
| DeepSeek-V4.1-Flash-1017 | 96%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
| OpenAI: GPT-6.1 Sol | 96%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
| Anthropic: Claude Sonnet 5.5 | 94.7%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
| OpenAI: GPT-6.1 Terra | 94.3%Higher is better | LiveBench Coding2.0 · 2026-10-25 (dataset 20261025) | Effort: max · single runIndependent resultEvaluation date not reported | Benchmark result ↗Checked |
Model plus agent examples
Terminal-Bench · 4.0; pass rates belong to the complete submitted setup
| Model | Pass rate | Agent configuration | Uncertainty and dates | Sources and dates |
|---|---|---|---|---|
| Opus 5.5 | 64.85%Terminal-Bench4.0 | claude-codeAgent version 2.1.284Reasoning effort: maxSource model ID: anthropic/claude-opus-5-5 | 95% CI 61.74–67.96%n = 330Evaluation date not reported | Agent submission ↗Checked |
| Sonnet 5.5 | 61.82%Terminal-Bench4.0 | claude-codeAgent version 2.1.284Reasoning effort: maxSource model ID: anthropic/claude-sonnet-5-5 | 95% CI 58.88–64.76%n = 330Evaluation date not reported | Agent submission ↗Checked |
| GPT-6.1 Sol | 58.18%Terminal-Bench4.0 | codexAgent version 0.159.0Reasoning effort: maxSource model ID: openai/gpt-6.1-sol | 95% CI 55.07–61.29%n = 330Evaluation date not reported | Agent submission ↗Checked |
| Fable 5.1 | 57.88%Terminal-Bench4.0 | claude-codeAgent version 2.1.257Reasoning effort: maxSource model ID: anthropic/claude-fable-5-1 | 95% CI 54.12–61.64%n = 330Evaluation date not reported | Agent submission ↗Checked |
| Opus 5 | 51.82%Terminal-Bench4.0 | claude-codeAgent version 2.1.231Reasoning effort: maxSource model ID: anthropic/claude-opus-5 | 95% CI 48.43–55.21%n = 330Evaluation date not reported | Agent submission ↗Checked |
| GPT-6 Sol | 49.39%Terminal-Bench4.0 | codexAgent version 0.159.0Reasoning effort: maxSource model ID: openai/gpt-6-sol | 95% CI 46.16–52.62%n = 330Evaluation date not reported | Agent submission ↗Checked |
Where these rows come from
We join catalog specifications, recorded provider quotes and published evaluation records where exact model identities are available. Some observations come through aggregators or benchmark collections. Multiple listings do not establish independent confirmation of a claim.
Open the source beside a row to check the original evidence. A collection or check date is separate from a model release or evaluation date. Unknown values stay unknown. Recorded prices are in USD with their original billing units and settings.
See data coverage and sources →Compare for your own use case
Keep your task, model identity and provider configuration together. Explore the catalog yourself, or connect your AI to read the same source-backed records.