FrontierCode — Results
| Rank | Model | Score |
|---|---|---|
| 1 | grok-4-6 | 61.3 |
FrontierCode
1 phaseActive
Cognition's rubric-weighted, blocking-criteria-gated coding agent benchmark with Main and Extended task subsets.
Overview
FrontierCode
Quick answer: FrontierCode is Cognition's coding agent benchmark that scores agentic coding runs using a rubric-weighted system gated by blocking criteria, rather than simple pass/fail — it publishes a 100-task Main subset and a 150-task Extended subset (a prior 50-task Diamond subset was deprecated in v1.1). Grok 4.6 scores 61.3% on the Extended subset as of August 2026.
At a Glance
What it tests: Whether a coding agent produces a working, correctly-scoped solution to a realistic engineering task, judged against a weighted rubric where certain "blocking" criteria must pass before partial credit counts.
Why it matters: The rubric-weighted, blocking-criteria design more closely mirrors how a human engineer reviews a pull request than a binary pass/fail test suite, making FrontierCode a signal for production-readiness rather than just raw problem-solving.
Known limitations: As a v1.1 benchmark from a single lab (Cognition), methodology transparency and independent replication are more limited than academic benchmarks like SWE-Bench.
What FrontierCode Measures
FrontierCode is developed by Cognition (maker of the Devin coding agent) to evaluate agentic coding models on realistic software engineering tasks. Unlike simple test-pass benchmarks, FrontierCode grades each submission against a weighted rubric: certain "blocking" criteria (e.g., the change must not break existing functionality) gate whether the submission can receive any credit at all, after which partial credit is awarded across secondary rubric items.
The benchmark ships in two published subsets — Main (100 tasks) and Extended (150 tasks) — with a legacy 50-task Diamond subset deprecated as of v1.1.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Coding agent |
| Metric | % score (rubric-weighted, blocking-criteria-gated) |
| Number of tasks | 100 (Main) / 150 (Extended) |
| Saturation | Low |
| Created by | Cognition |
| Project | cognition.com/frontiercode |
How FrontierCode Is Scored
Each task submission is graded against a rubric with blocking criteria that must be satisfied before any partial credit is awarded; the final score is a weighted percentage across all rubric items for all tasks in the chosen subset (Main or Extended).
State-of-the-Art Results
| Rank | Model | Score | Subset | Source | Date |
|---|---|---|---|---|---|
| 1 | Grok 4.6 | 61.3% | Extended | xAI Grok 4.6 announcement | 2026-08 |
Score sourced from xAI's Grok 4.6 announcement, August 2026, reporting the Extended subset. Cognition's own leaderboard reports Main-subset scores separately at cognition.com/frontiercode; scores across subsets are not directly comparable.
FrontierCode on Benchgen
No Benchgen results yet — be the first to run FrontierCode.
FrontierCode vs Other Benchmarks
| Benchmark | What it tests | Tasks | Saturation |
|---|---|---|---|
| FrontierCode | Rubric-weighted, blocking-gated coding agent tasks | 100–150 | Low |
| DeepSWE | Real-world software engineering agent tasks | — | Low |
| APEX-SWE | Pass@1 software engineering agent tasks | 200 | Low |
Run FrontierCode on Your Model
Benchgen lets teams run FrontierCode-style rubric evaluations against their own model and harness combinations, tracking regressions across versions rather than relying on a single vendor-reported number.