Terminal-Bench 4.0

Every result

• Claude Sonnet 5.570.6%Reported by the model's maker (Anthropic), Sep 28, 2026

  Anthropic reports 70.6% accuracy for Claude Sonnet 5.5 on Terminal-Bench 4.0. Benchmark effort and full evaluation methods remain unknown.

  Limits: Provider report only; no independent run, private test, overall ranking or matched-budget comparison. The announcement does not identify Sonnet effort for this result. API high and app Medium defaults do not establish the benchmark setting. Exact task subset, sample size, trials/pass@k, harness, prompts, tool/reasoning budgets, aggregation, execution date and uncertainty remain unknown. Announcement describes higher-risk cybersecurity fallback to Sonnet 5 at launch; whether safeguards were enabled or any fallback triggered on this benchmark is unknown. No isolated fixed-weight claim. Full evaluation methods remain unverified; the announcement gives only partial methodological detail. Opus 5.5 has a separately reported 66.4% Terminal-Bench 4.0 result at xhigh with production safeguards and possible fallbacks. Shared benchmark/version does not establish matched methods or a meaningful 4.2-point advantage.

  Anthropic · Anthropic · Anthropic · Anthropic

• Claude Opus 5.566.4%Reported by the model's maker (Anthropic), Sep 22, 2026

  Anthropic reports 66.4% Terminal-Bench 4.0 accuracy for Claude Opus 5.5 at xhigh effort. The evaluated serving configuration used production safeguards and could route some tasks to fallback models.

  Setting: Adaptive thinking at xhigh effort for the reported Terminal-Bench 4.0 row; numerical reasoning/token budget unknown.

  Limits: Provider-reported result, not an independent run, private test, overall ranking or matched-budget comparison. The reported ±2.6 percentage points is standard error, not a confidence interval. The documented serving configuration can include Opus 4.8 cybersecurity and Opus 5 biology/frontier-development fallbacks. Whether any fallback actually ran on this Terminal-Bench row is unestablished; no pure-weights claim. Exact task subset, sample size, trial count, pass@k, aggregation, harness, prompts, tool budget and execution date remain unknown. Public leaderboard trial counts for other models do not establish this run's settings. Full evaluation methods remain unverified; the announcement gives only partial methodological detail. The API default medium effort differs from this xhigh benchmark setting. No benchmark task-cost or equal-effort comparison is established.

  Anthropic · Anthropic · Anthropic · Anthropic

• Claude Opus 552.3%Reported by the model's maker (Anthropic), Sep 22, 2026

  Anthropic reports 52.3% Terminal-Bench 4.0 accuracy for Claude Opus 5 in its own reproduction. The evaluated checkpoint and full configuration remain unknown.

  Limits: Provider-reported result tied to an older named comparison model, not an independently verified run, overall ranking or matched-budget comparison. This is a named model comparison, not a specific evaluated checkpoint. Its API identifier, build and serving configuration are unknown. Effort, guardrails, task subset, sample/trial count, harness, prompts, tool/reasoning/token budgets, aggregation, pass-k and execution date remain unknown. The newer 5.5 models' defaults do not establish this older result's settings. Anthropic distinguishes its 52.3% reproduction from a public leaderboard's 51.8%. The public five-trials/task and Claude Code harness conditions do not establish the reproduction's protocol. The reported standard-error range for other Claude models does not identify this row's uncertainty. No per-model confidence interval was established. Opus 5.5's ±2.6 percentage-point standard error describes another result. Shared Terminal-Bench 4.0 naming and percent units with current 5.5 reports do not establish matched artifacts, protocols, budgets or meaningful score-gap comparisons.

  Anthropic · Anthropic

• Gemini 3.8 Flash19.1%Reported by the model's maker (Google DeepMind)

  Google DeepMind reports 19.1% on Terminal-Bench 4.0 for Gemini 3.8 Flash, with results described as of September 2026. The exact evaluated build and full configuration remain unknown.

  Limits: Provider-reported own-model result; the actual execution runner and scorer are not identified. No independent replication, overall ranking or matched-budget comparison is established. The source identifies the nominal Gemini 3.8 Flash release, not an exact evaluated checkpoint. Equivalence to the currently served API build and its routing or fallback configuration is unknown. Task subset, sample size, trial count, harness, prompting, tool mode and budget, reasoning effort, token budget, safeguards, aggregation, pass-k and uncertainty are not established by the inspected HTML. Medium is the API product default; it is not established as the effort used for this result. Settings from other benchmarks and models are not transferred. The source page is dated September 2, 2026 and describes results as of September 2026. The exact evaluation/run date and exact date on which this scalar was added are unknown. A shared benchmark name and percentage unit with another provider’s report do not establish matched methods or justify score-gap claims.

  Google DeepMind · Google

• Claude Sonnet 510.3%Reported by the model's maker (Anthropic), Sep 28, 2026

  Anthropic reports 10.3% Terminal-Bench 4.0 accuracy for its older Claude Sonnet 5 model. The evaluated checkpoint and full configuration remain unknown.

  Limits: Provider-reported result tied to an older named comparison model, not an independently verified run, overall ranking or matched-budget comparison. This is a named model comparison, not a specific evaluated checkpoint. Its API identifier, build and serving configuration are unknown. Effort, guardrails, task subset, sample/trial count, harness, prompts, tool/reasoning/token budgets, aggregation, pass-k and execution date remain unknown. The newer 5.5 models' defaults do not establish this older result's settings. No per-model confidence interval was established. Opus 5.5's ±2.6 percentage-point standard error describes another result. Shared Terminal-Bench 4.0 naming and percent units with current 5.5 reports do not establish matched artifacts, protocols, budgets or meaningful score-gap comparisons.

  Anthropic · Anthropic