FrontierCode v1.1 Main Leaderboard
Production-code quality benchmark by Cognition. 100 private tasks scored by maintainer-authored rubrics — correctness, tests, scope, style, and maintainability. Models ranked by published Main scores.
Updated: October 30, 2026 · FrontierCode 1.1 is now an archived benchmark; FrontierCode 2.0 is the current suite. Adds an independent Kimi K3.1 Pro 1.1 Main result (47.1%, Artificial Analysis Stirrup; not a Cognition rank). See the interactive FrontierCode 2.0 leaderboard → · Historical source: Cognition · SWE-bench Pro → · Terminal-Bench →
#
Model
Prov
Main
Harness
License
$/1M Out
Source
Src
FrontierCode v1.1 Main Score Comparison
Source-specific Main scores on a 0–100 scale. Independent Artificial Analysis runs are shown in the chart but carry no Cognition rank; compare across harnesses with caution.
Version status (Oct 30, 2026): FrontierCode v1.1 is now archived; Cognition's live board moved to v2.0 on Oct 27. The task set and evaluation changed, so v2.0 scores are not comparable to this v1.1 Main table. See the current FrontierCode 2.0 interactive leaderboard.
About FrontierCode v1.1: Cognition's production-code benchmark. Tests whether AI-generated pull requests are mergeable — scored for correctness, tests, scope, style, and maintainability through maintainer-authored rubrics. 100 private Main tasks (150 in Extended). Each row combines a model with an agent harness at its best-performing effort.
🆕 Cognition board refresh (checked Oct 23; latest additions: DeepSeek V4.1 Flash and Mistral Large 4 on Oct 13): Claude Opus 5.5 leads FrontierCode 1.1 Main at 54.6% (medium, $0.80 mean cost per rollout, 18.5k tokens). Sonnet 5.5 scores 52.1% (xhigh, $1.59/run, 47.4k tokens); GPT-6.1 Sol is 50.2% (medium, $0.36/run, 11.4k tokens); Haiku 5.5 posts 46.4% (max, $1.33/run, 181.4k tokens). Other board rows include SWE-2 (50.0%, $1.18/run), GPT-6 Sol (49.3%, $2.07/run), and Grok 4.7 (47.6%, $6.65/run). Rollout spend is distinct from API token rates. Haiku API price is $0.10/$0.50 per 1M input/output for prompts up to 100K tokens ($0.50/$2.50 above that threshold).
Anthropic’s separate launch report: Opus 5.5 scores 54.4% at max effort on its own FrontierCode run. Keep that provider run distinct from Cognition’s 54.6% medium-effort leaderboard configuration.
Separate Extended-set result (not ranked in this Main table): Cognition’s Devin release note reports GPT-6.1 Sol at 60.4% on FrontierCode 1.1 Extended, in line with GPT-6 Sol (60.7%). GPT-6.1 Sol’s medium-effort rollout cost was $0.31; Extended scores are not comparable to the 100-task Main scores above. Source: Devin / Cognition.
🆕 GPT-5.6 Sol (47.5%): OpenAI's flagship scores 47.5% Main on the official Cognition leaderboard. Extended: 60.6%.
🆕 Grok 4.6 (48.0%): Ranks #3 on the official Cognition leaderboard — ahead of GPT-5.6 Sol — at high effort ($2.88/run). Gemini 3.7 Flash (43.6%) is the strongest Flash-tier result yet, +9.2 pts over 3.6 Flash, at medium effort ($3.65/run). Both landed the week of Aug 12–13, 2026.
🆕 New entries (Aug 2026): Kimi K3 (44.2%), Grok 4.5 (42.4%), GPT-5.6 Terra (41.3%), GPT-5.6 Luna (39.8%), Gemini 3.6 Flash (34.4%), DeepSeek V4 Flash 0731 (18.8%), DeepSeek V4 Pro (17.6%), MiniMax M3 (14.7%), Inkling 0.99 (14.0%), Mistral 3.5 Medium (8.0%), and Qwen 3.7 Plus (10.2%) all now have published Main scores on the official Cognition FrontierCode leaderboard.
🆕 Independent Artificial Analysis snapshot (Oct 12): Cognition’s Oct 13 additions (DeepSeek V4.1 Flash 43.9%, Mistral Large 4 Preview 39.4%) appear above as separate official Cognition rows with rollout costs. The Cognition board’s Oct 23 changelog has no later additions in the visible source. AA re-ran FrontierCode 1.1 Main in its Stirrup harness (100 private tasks, three attempts per task, GPT-6 Luna judge), reporting Opus 5.5 52.2% ($10.15/task), GPT-6.1 Sol 49.1% ($0.81/task), Fable 5.1 49.0% ($12/task). The current AA page also lists Opus 5.5 at 57.1% (medium, default fallback); this live setting/result is separately labeled and its run cost is not displayed in the page summary. Other AA launch-snapshot results include MiMo-V2.6-Pro 45% ($0.18/task), DeepSeek V4.1 Flash and Kimi K3 44% each, and GPT-6 Luna 42% ($0.09/task). These are separate independent runs, not Cognition's leaderboard scores; they have no Cognition rank. Artificial Analysis results & methodology.
Harness note: Scores combine different agent harnesses (Cognition configurations and independent Stirrup). Results from different harnesses are directional rather than directly comparable.
About FrontierCode v1.1: Cognition's production-code benchmark. Tests whether AI-generated pull requests are mergeable — scored for correctness, tests, scope, style, and maintainability through maintainer-authored rubrics. 100 private Main tasks (150 in Extended). Each row combines a model with an agent harness at its best-performing effort.
🆕 Cognition board refresh (checked Oct 23; latest additions: DeepSeek V4.1 Flash and Mistral Large 4 on Oct 13): Claude Opus 5.5 leads FrontierCode 1.1 Main at 54.6% (medium, $0.80 mean cost per rollout, 18.5k tokens). Sonnet 5.5 scores 52.1% (xhigh, $1.59/run, 47.4k tokens); GPT-6.1 Sol is 50.2% (medium, $0.36/run, 11.4k tokens); Haiku 5.5 posts 46.4% (max, $1.33/run, 181.4k tokens). Other board rows include SWE-2 (50.0%, $1.18/run), GPT-6 Sol (49.3%, $2.07/run), and Grok 4.7 (47.6%, $6.65/run). Rollout spend is distinct from API token rates. Haiku API price is $0.10/$0.50 per 1M input/output for prompts up to 100K tokens ($0.50/$2.50 above that threshold).
Anthropic’s separate launch report: Opus 5.5 scores 54.4% at max effort on its own FrontierCode run. Keep that provider run distinct from Cognition’s 54.6% medium-effort leaderboard configuration.
Separate Extended-set result (not ranked in this Main table): Cognition’s Devin release note reports GPT-6.1 Sol at 60.4% on FrontierCode 1.1 Extended, in line with GPT-6 Sol (60.7%). GPT-6.1 Sol’s medium-effort rollout cost was $0.31; Extended scores are not comparable to the 100-task Main scores above. Source: Devin / Cognition.
🆕 GPT-5.6 Sol (47.5%): OpenAI's flagship scores 47.5% Main on the official Cognition leaderboard. Extended: 60.6%.
🆕 Grok 4.6 (48.0%): Ranks #3 on the official Cognition leaderboard — ahead of GPT-5.6 Sol — at high effort ($2.88/run). Gemini 3.7 Flash (43.6%) is the strongest Flash-tier result yet, +9.2 pts over 3.6 Flash, at medium effort ($3.65/run). Both landed the week of Aug 12–13, 2026.
🆕 New entries (Aug 2026): Kimi K3 (44.2%), Grok 4.5 (42.4%), GPT-5.6 Terra (41.3%), GPT-5.6 Luna (39.8%), Gemini 3.6 Flash (34.4%), DeepSeek V4 Flash 0731 (18.8%), DeepSeek V4 Pro (17.6%), MiniMax M3 (14.7%), Inkling 0.99 (14.0%), Mistral 3.5 Medium (8.0%), and Qwen 3.7 Plus (10.2%) all now have published Main scores on the official Cognition FrontierCode leaderboard.
🆕 Independent Artificial Analysis snapshot (Oct 12): Cognition’s Oct 13 additions (DeepSeek V4.1 Flash 43.9%, Mistral Large 4 Preview 39.4%) appear above as separate official Cognition rows with rollout costs. The Cognition board’s Oct 23 changelog has no later additions in the visible source. AA re-ran FrontierCode 1.1 Main in its Stirrup harness (100 private tasks, three attempts per task, GPT-6 Luna judge), reporting Opus 5.5 52.2% ($10.15/task), GPT-6.1 Sol 49.1% ($0.81/task), Fable 5.1 49.0% ($12/task). The current AA page also lists Opus 5.5 at 57.1% (medium, default fallback); this live setting/result is separately labeled and its run cost is not displayed in the page summary. Other AA launch-snapshot results include MiMo-V2.6-Pro 45% ($0.18/task), DeepSeek V4.1 Flash and Kimi K3 44% each, and GPT-6 Luna 42% ($0.09/task). These are separate independent runs, not Cognition's leaderboard scores; they have no Cognition rank. Artificial Analysis results & methodology.
Harness note: Scores combine different agent harnesses (Cognition configurations and independent Stirrup). Results from different harnesses are directional rather than directly comparable.
📊 See also: SWE-bench Pro Leaderboard — real-world bug fixing | Terminal-Bench 2.1 Leaderboard — CLI coding | DeepSWE v1.1 Leaderboard — long-horizon engineering
Test these models on production code
20+ LLMs on CodingFleet. Run them on your own PRs. Benchmarks are a compass, not a map.
🚀 Try on CodingFleet →