LLM Reference
activeCodingAgents

Frontier-Bench v0.1

Metric: Mean reward (higher is better)Introduced: 2026

A 74-task benchmark across seven professional domains for long-horizon autonomous work. Anthropic's Opus 5 system-card run uses the mini-SWE-agent harness, a Google Kubernetes Engine backend, and mean reward over five attempts per task.

Models ranked

1

tracked on this benchmark

Score band

44.4 – 44.4

best → lowest tracked

Snapshot trend

—

need ≥2 snapshots

Leaderboard

Tracked models ranked by Mean reward (higher is better).

Compare candidates
#Model variant and provenanceScore
1
Claude Opus 5

Configuration: xhigh effort

Version: Frontier-Bench v0.1Harness: 74 tasks; mini-SWE-agent on Google Kubernetes Engine; mean reward over five attempts per taskEvaluator: AnthropicObserved: Jul 24, 2026Confidence: confirmedSource

Notes: Vendor-reported. Source: p.152, Figure 8.5. About 5% of API calls were refused and 4% of trials fell back to Opus 4.8; maximum effort scored 43.0 within noise of xhigh.

44.4

How to read this benchmark

This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.

Trust this score when

  • There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
  • The model list covers the same version family you can actually deploy today.
  • Top candidates overlap with your required routing and feature requirements.

Be cautious when

  • There is only one benchmark snapshot or the dataset appears stale.
  • The benchmark metric direction is opposite of your decision objective.
  • The score difference between options is narrow and likely within implementation variance.

Related benchmarks

HumanEval
MBPP
Mostly Basic Programming Problems
HumanEval+
MBPP+
Mostly Basic Programming Problems+
Spider
ToolTalk

Last reviewed: Jul 26, 2026

Resources

WebsiteLeaderboard