Skip to main content
BenchLM
Data

FrontierCode 1.1 Main

We show this table for reference; we do not rank on it.

Cognition's 100-task software-engineering benchmark for whether coding agents produce mergeable, production-quality pull requests, scored for correctness, tests, scope, style, and maintainability through maintainer-authored rubrics.

Current release

FrontierCode 2.0 is the current release. Its 132 tasks are larger, come from more repositories, and add stricter blocking criteria, especially for thoroughness, so its scores are not directly comparable with FrontierCode 1.1.

View FrontierCode 2.0

Main score on FrontierCode 1.1 Main — November 12, 2026 snapshot

We mirror the published main score view for FrontierCode 1.1 Main. Claude Opus 5.5 leads the public snapshot at 54.6%, followed by Claude Fable 5 (53.5%) and Claude Opus 5 (53.4%). We do not use these results to rank models overall.

1Closed

Claude Opus 5.5

Anthropic

claude-code · medium reasoning

54.6%
Overall 86.62Context 1M
2Closed

Claude Fable 5

Anthropic

claude-code · xhigh reasoning

53.5%
Overall 76.28Context 1M+
3Closed

Claude Opus 5

Anthropic

claude-code · medium reasoning

53.4%
Overall 76.05Context 1M

45 modelsCodingCurrentDisplay onlyUpdated November 12, 2026 snapshot

Main score table (45 models)

Main score results for FrontierCode 1.1 Main
RankModel / configurationScoreParameters (B)Open / closed
1Claude Opus 5.5Anthropicclaude-code · medium reasoning
54.6%
Not reportedClosed
2Claude Fable 5Anthropicclaude-code · xhigh reasoning
53.5%
Not reportedClosed
3Claude Opus 5Anthropicclaude-code · medium reasoning
53.4%
Not reportedClosed
4GPT-6 AstraOpenAIcodex · max reasoning
53.3%
Not reportedClosed
5Claude Sonnet 5.5Anthropicclaude-code · xhigh reasoning
52.1%
Not reportedClosed
6Claude Fable 5.1Anthropicclaude-code · medium reasoning
50.9%
Not reportedClosed
7GPT-6.1 SolOpenAIcodex · medium reasoning
50.2%
Not reportedClosed
8SWE-2Cognitiondevin · max reasoning
50.0%
Not reportedClosed
9GPT-6 SolOpenAIcodex · max reasoning
49.3%
Not reportedClosed
10Grok 4.6xAIgrok-build · high reasoning
48.0%
Not reportedClosed
11Grok 4.7xAIgrok-build · high reasoning
47.6%
Not reportedClosed
12Gemini 4 ArgonGoogledevin-cli · medium reasoning
47.5%
Not reportedClosed
13GPT-5.6 SolOpenAIcodex · max reasoning
47.5%
Not reportedClosed
14Claude Opus 4.8Anthropicclaude-code · max reasoning
46.5%
Not reportedClosed
15Claude Haiku 5.5Anthropicclaude-code · max reasoning
46.4%
Not reportedClosed
16Kimi K3Moonshot AImini-swe-agent · none reasoning
44.2%
Not reportedOpen
17DeepSeek V4.1 Flash 1017DeepSeekdevin-cli · high reasoning
43.9%
Not reportedOpen
18Gemini 3.7 FlashGoogledevin-cli · medium reasoning
43.6%
Not reportedClosed
19GPT-5.5OpenAIcodex · xhigh reasoning
43.0%
Not reportedClosed
20Claude Sonnet 5Anthropicclaude-code · xhigh reasoning
42.7%
Not reportedClosed
21GPT-6 LunaOpenAIcodex · max reasoning
42.4%
Not reportedClosed
22Grok 4.5xAIgrok-build · high reasoning
42.4%
Not reportedClosed
23SWE-1.7Cognitiondevin-cli · none reasoning
42.0%
Not reportedClosed
24GPT-5.6 TerraOpenAIcodex · max reasoning
41.3%
Not reportedClosed
25Gemini 3.8 FlashGoogledevin-cli · medium reasoning
41.2%
Not reportedClosed
26GLM-5.3Z.AIdevin-cli · max reasoning
40.1%
Not reportedOpen
27GPT-5.6 LunaOpenAIcodex · max reasoning
39.8%
Not reportedClosed
28Claude Opus 4.7Anthropicclaude-code · max reasoning
38.5%
Not reportedClosed
29Gemini 3.6 FlashGoogledevin-cli · medium reasoning
34.4%
Not reportedClosed
30GLM-5.3-FlashZ.AIdevin-cli · max reasoning
31.8%
Not reportedOpen
31Kimi K2.7Moonshot AImini-swe-agent · none reasoning
30.1%
Not reportedUnknown
32DeepSeek V4 Pro 0813DeepSeekdevin-cli · high reasoning
28.5%
Not reportedOpen
33GPT-5.4 miniOpenAIcodex · xhigh reasoning
27.0%
Not reportedClosed
34Claude Opus 4.6Anthropicclaude-code · high reasoning
26.6%
Not reportedClosed
35Composer 2.5Cursorcursor-cli · none reasoning
25.6%
Not reportedClosed
36GLM-5.2Z.AImini-swe-agent · none reasoning
24.5%
Not reportedOpen
37Claude Sonnet 4.6Anthropicclaude-code · max reasoning
24.3%
Not reportedClosed
38DeepSeek V4 Flash 0731DeepSeekdevin-cli · high reasoning
18.8%
Not reportedOpen
39DeepSeek V4 ProDeepSeekmini-swe-agent · none reasoning
17.6%
Not reportedUnknown
40MiniMax M3MiniMaxmini-swe-agent · none reasoning
14.7%
Not reportedOpen
41InklingThinking Machines Labmini-swe-agent · 0.99 reasoning
14.0%
Not reportedOpen
42Nemotron 3 UltraNVIDIAdevin-cli · none reasoning
13.6%
Not reportedOpen
43Qwen3.7 PlusAlibabamini-swe-agent · none reasoning
10.2%
Not reportedClosed
44SWE-1.6Cognitiondevin-cli · none reasoning
9.4%
Not reportedUnknown
45Mistral 3.5 MediumMistraldevin-cli · none reasoning
8.0%
Not reportedUnknown

Results reported by model developersUnranked

Labs ran these themselves. They take no rank here.

FrontierCode 1.1 Main results reported by model developers, unranked
Claude Haiku 5.5Anthropic45.8%Task scoreFrontierCode 1.1 (Main)xhigh effort; the stored score is the max-effort runSame measure as the board score, run by the labReported by Anthropic
Claude Haiku 5.5 system card

A lab’s own result is shown for reference. It never takes a rank on this board.

How we label sources

How we show FrontierCode 1.1 Main

The snapshot mirrors Cognition's FrontierCode 1.1 Main table captured on November 12, 2026 snapshot. Main uses the 100 hardest tasks from the 150-task Extended set. Each table selects the best published reasoning effort for its own task set.

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. Runs flagged for unfair internet use receive zero in the published score.

Snapshot

45 model-agent rows100 Main tasks5 trials per effortBest-performing effortDisplay only
FrontierCode leaderboardFrontierCode 1.1 methodologyOfficial result data JSONExtended: 150 tasks

The published FrontierCode 1.1 Main snapshot places Claude Opus 5.5 first at 54.6%. The third row is 1.2 points behind. The broader top-10 range is 6.6 points, so many of the published results sit in a relatively narrow band.

45 models have been evaluated on FrontierCode 1.1 Main. The benchmark falls in the Coding category. FrontierCode 1.1 Main is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings. We moved it to display only on November 12, 2026 because FrontierCode 2.0 supersedes version 1.1.

About FrontierCode 1.1 Main

Year

2026

Tasks

100 private Main tasks (150 in Extended)

Format

Repository task completion with maintainer rubrics

Difficulty

Frontier coding-agent quality

FrontierCode 1.1 Main uses 100 of the benchmark's 150 private software-engineering tasks. The leaderboard reports the best-performing published reasoning effort for each model-agent row. Each row combines a model with an agent harness, and the private tasks cannot be independently rerun from the public artifact.

FrontierCode leaderboardPublic benchmark source

Freshness and provenance

Version

FrontierCode 1.1 Main

Refresh cadence

Rolling

Staleness state

Current

Question availability

Private tasks with public aggregate results

CurrentDisplay only

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 1.1 Main measure?

Cognition's 100-task software-engineering benchmark for whether coding agents produce mergeable, production-quality pull requests, scored for correctness, tests, scope, style, and maintainability through maintainer-authored rubrics.

Which model leads the published FrontierCode 1.1 Main snapshot?

Claude Opus 5.5 currently leads the published FrontierCode 1.1 Main snapshot with 54.6% main score. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on FrontierCode 1.1 Main?

The November 12, 2026 snapshot snapshot contains 45 AI models.

Last updated: November 12, 2026 snapshot · mirrored from the public benchmark leaderboard

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.