FrontierCode 2.0
We show this table for reference; we do not rank on it.
Cognition's second-generation coding-agent benchmark, scoring whether agents produce mergeable pull requests on larger, less fully specified tasks, with added emphasis on thoroughness such as fixing every instance of a bug.
Previous release
FrontierCode 1.1 Main and Extended remain available as separate tables. FrontierCode 2.0 uses a new task set and stricter grading, so a lower 2.0 score does not mean a model got worse.
View FrontierCode 1.1Score on FrontierCode 2.0 — October 27, 2026 snapshot
We mirror the published score view for FrontierCode 2.0. Claude Opus 5.5 leads the public snapshot at 45.5%, followed by GPT-6 Astra (42.8%) and Claude Sonnet 5.5 (41.0%). We do not use these results to rank models overall.
Claude Opus 5.5
Anthropic
claude-code · max reasoning
GPT-6 Astra
OpenAI
codex · max reasoning
Claude Sonnet 5.5
Anthropic
claude-code · max reasoning
21 modelsCodingCurrentDisplay onlyUpdated October 27, 2026 snapshot
Score table (21 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Claude Opus 5.5Anthropicclaude-code · max reasoning | 45.5% | Not reported | Closed |
| 2 | GPT-6 AstraOpenAIcodex · max reasoning | 42.8% | Not reported | Closed |
| 3 | Claude Sonnet 5.5Anthropicclaude-code · max reasoning | 41.0% | Not reported | Closed |
| 4 | GPT-6.1 SolOpenAIcodex · max reasoning | 39.0% | Not reported | Closed |
| 5 | GPT-6 SolOpenAIcodex · max reasoning | 34.3% | Not reported | Closed |
| 6 | Claude Fable 5.1Anthropicclaude-code · xhigh reasoning | 32.4% | Not reported | Closed |
| 7 | Grok 4.7xAIgrok-build · xhigh reasoning | 28.6% | Not reported | Closed |
| 8 | Claude Haiku 5.5Anthropicclaude-code · max reasoning | 27.7% | Not reported | Closed |
| 9 | GPT-5.6 SolOpenAIcodex · max reasoning | 27.6% | Not reported | Closed |
| 10 | Claude Opus 5Anthropicclaude-code · xhigh reasoning | 26.7% | Not reported | Closed |
| 11 | Gemini 4 ArgonGooglechisel · high reasoning | 26.3% | Not reported | Closed |
| 12 | Muse WonderMetamuse · high reasoning | 24.3% | Not reported | Closed |
| 13 | SWE-2Cognitiondevin · max reasoning | 23.8% | Not reported | Closed |
| 14 | GPT-5.5OpenAIcodex · xhigh reasoning | 22.6% | Not reported | Closed |
| 15 | GPT-6 LunaOpenAIcodex · max reasoning | 21.9% | Not reported | Closed |
| 16 | DeepSeek V4.1 Flash 1017DeepSeekchisel · max reasoning | 19.3% | Not reported | Open |
| 17 | Kimi K3Moonshot AIchisel · high reasoning | 17.7% | Not reported | Pending |
| 18 | SWE-1.7Cognitionchisel_swe1 | 16.2% | Not reported | Closed |
| 19 | Claude Opus 4.8Anthropicclaude-code · max reasoning | 14.7% | Not reported | Closed |
| 20 | Mistral Large 4Mistralchisel · high reasoning | 8.8% | Not reported | Pending |
| 21 | SWE-1.6Cognitionchisel_swe1 | 1.3% | Not reported | Unknown |
How we show FrontierCode 2.0
The snapshot mirrors the 21 model-agent rows in Cognition's FrontierCode 2.0 launch post, captured on October 27, 2026, across 132 private tasks. Each row uses the model's best published reasoning effort, and a solution that fails any blocking criterion scores zero.
We keep these mirrored results display-only. Each row combines a model with an agent harness, and the tasks are private, so the mirror does not enter weighted model rankings. FrontierCode 2.0 uses a new task set and stricter grading, so its scores are not comparable with FrontierCode 1.1.
Snapshot
The published FrontierCode 2.0 snapshot places Claude Opus 5.5 first at 45.5%. The third row is 4.5 points behind. The broader top-10 range is 18.8 points, so the table still separates the published systems.
21 models have been evaluated on FrontierCode 2.0. The benchmark falls in the Coding category. FrontierCode 2.0 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About FrontierCode 2.0
Year
2026
Tasks
132 private software-engineering tasks
Format
Repository task completion with maintainer rubrics
Difficulty
Frontier coding-agent quality
FrontierCode 2.0 has 132 private tasks built by open-source maintainers from their own repositories. Compared with FrontierCode 1.1, tasks come from 55% more repositories, carry 43% more blocking criteria, and their reference solutions change 67% more files and 89% more lines. The score is a weighted aggregate of rubric items, and a solution that fails any blocking criterion scores zero. The mirrored table uses each model-agent row's best published reasoning effort; scores are not comparable with FrontierCode 1.1.
Freshness and provenance
Version
FrontierCode 2.0
Refresh cadence
Rolling
Staleness state
Current
Question availability
Private tasks with public aggregate results
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does FrontierCode 2.0 measure?
Cognition's second-generation coding-agent benchmark, scoring whether agents produce mergeable pull requests on larger, less fully specified tasks, with added emphasis on thoroughness such as fixing every instance of a bug.
Which model leads the published FrontierCode 2.0 snapshot?
Claude Opus 5.5 currently leads the published FrontierCode 2.0 snapshot with 45.5% score. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on FrontierCode 2.0?
The October 27, 2026 snapshot snapshot contains 21 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 6,100+ subscribers.
One email each week. Unsubscribe anytime.