- Context
- 1.1Mtokens
- Released
- 4mo ago
Cost/1M tokens
Benchmarks
DeepSWE is the primary coding benchmark where an exact model evaluation exists. Arena leaderboards and other sourced evaluations appear below; missing scores are not zero.
How it ranks
Against its predecessor, similarly priced models and the models ranked around it.
DeepSWE 1.1
Best pass@1 per model · Higher is better
#6 of 28
- 74.1%
- 73.8%
- 73.6%
- 72.7%
- 69.9%
- 69.6%GPT-5.6 Terra
- 69.0%
- 68.5%
- 67.5%
- 67.2%
Arena Text
Best Arena rating per model · Higher is better
Arena #51
- 1487Kimi K3Similar price
- 1487Gemini 3.1 Pro PreviewSimilar price
- 1481GPT-5.5Predecessor
- 1472Claude Sonnet 5.5Similar price
- 1472Claude Sonnet 4.6Nearby rank
- 1468GPT-5.6 Terra
- 1467
ERNIE 5.1Nearby rank - 1467
MiMo-V2.5-ProNearby rank - 1354
Command ASimilar price
-13 vs GPT-5.5 (predecessor) · Ahead of 1 of 4 similarly priced models
Arena WebDev
Best Arena rating per model · Higher is better
Arena #61
- 1773Claude Sonnet 5.5Similar price
- 1676
Qwen3.8 Max (0902)Similar price - 1658Kimi K3Similar price
- 1526
Muse Spark OpenNearby rank - 1522Claude Sonnet 4.6Nearby rank
- 1520GPT-5.6 Terra
- 1519GPT-5.6 LunaNearby rank
- 1515Qwen3.7 MaxNearby rank
- 1512GPT-5.5Predecessor
- 1446Gemini 3.1 Pro PreviewSimilar price
Level with GPT-5.5 (predecessor, overlapping 95% CIs) · Ahead of 1 of 4 similarly priced models
Arena Vision
Best Arena rating per model · Higher is better
Arena #37
- 1285GPT-5.5Predecessor
- 1280Grok 4.5Similar price
- 1280Gemini 3.1 Pro PreviewSimilar price
- 1275Claude Sonnet 4.6Nearby rank
- 1272GLM 5.3 FlashNearby rank
- 1266GPT-5.6 Terra
- 1266Gemini 3.5 Flash LiteNearby rank
- 1265Gemma 4 31BNearby rank
- 1263Claude Sonnet 5.5Similar price
- 1199
Mistral Medium 3.5Similar price
-18 vs GPT-5.5 (predecessor) · Ahead of 1 of 4 similarly priced models
Arena Search
Best Arena rating per model · Higher is better
Not ranked on Arena Search yet
Arena Document
Best Arena rating per model · Higher is better
Arena #15
- 1481GPT-5.5Predecessor
- 1480Muse Spark 1.1Nearby rank
- 1478Claude Sonnet 4.6Nearby rank
- 1474GPT-5.6 Terra
- 1474Muse Spark 1.3Nearby rank
- 1474GPT-6 AstraNearby rank
- 1469Claude Sonnet 5Similar price
- 1466Grok 4.5Similar price
- 1459Gemini 3.1 Pro PreviewSimilar price
- 1433Gemini 2.5 ProSimilar price
Level with GPT-5.5 (predecessor, overlapping 95% CIs) · Ahead of 2 of 4 similarly priced models
Agent Arena
Best IPS score per model · Higher is better
Arena #35
- +0.160Claude Sonnet 5.5Similar price
- +0.026Grok 4.7Similar price
- +0.006Kimi K3Similar price
- -0.000MiMo-V2.6-FlashNearby rank
- -0.002Grok 4.6Nearby rank
- -0.018GPT-5.6 Terra
- -0.022Gemini 3.7 FlashNearby rank
- -0.027
MiniMax M3.1 FlashNearby rank - -0.045Gemini 3.1 Pro PreviewSimilar price
Ahead of 1 of 4 similarly priced models
ARC-AGI-2
Best score per model · Higher is better
#16 of 85
- 85.0%GPT-5.5Predecessor
- 84.6%Gemini 3.7 FlashNearby rank
- 84.6%GPT-5.5 ProNearby rank
- 83.9%GPT-5.6 Terra
- 83.9%Claude Sonnet 5.5Similar price
- 83.3%GPT-5.4 ProNearby rank
- 77.1%Gemini 3.1 Pro PreviewSimilar price
- 67.1%Grok 4.6Similar price
- 60.4%Kimi K3Similar price
-1.1 pts vs GPT-5.5 (predecessor) · Ahead of 3 of 4 similarly priced models
FrontierMath-Tier-4-v2-Private
Best score per model · Higher is better
#17 of 72
- 80.5%Claude Sonnet 5.5Similar price
- 73.2%Claude Opus 5Nearby rank
- 72.5%GPT-5.5Predecessor
- 70.7%GPT-5.6 Terra
- 61.0%GPT-5.6 LunaNearby rank
- 58.5%GPT-5.4 ProNearby rank
- 39.0%Kimi K3Similar price
- 34.1%Qwen3.8 Max (0902)Similar price
- 26.8%Gemini 3.1 Pro PreviewSimilar price
-1.8 pts vs GPT-5.5 (predecessor) · Ahead of 3 of 4 similarly priced models
FrontierMath-Tiers-1-3-v2-Private
Best score per model · Higher is better
#13 of 87
- 88.8%Claude Sonnet 5.5Similar price
- 87.0%Claude Fable 5Nearby rank
- 86.0%Gemini 4 ArgonNearby rank
- 86.0%GPT-5.6 Terra
- 85.6%Claude Opus 5Nearby rank
- 85.3%GPT-5.5Predecessor
- 72.2%Kimi K3Similar price
- 66.0%Grok 4.6Similar price
- 59.6%Gemini 3.1 Pro PreviewSimilar price
+0.7 pts vs GPT-5.5 (predecessor) · Ahead of 3 of 4 similarly priced models
GPQA diamond
Best score per model · Higher is better
#18 of 161
- 95.6%Claude Sonnet 5.5Similar price
- 94.4%Gemini 3.1 Pro PreviewSimilar price
- 94.0%Grok 4.6Similar price
- 93.5%GPT-5.6 SolNearby rank
- 93.4%Grok 4.5Nearby rank
- 93.3%GPT-5.6 Terra
- 93.3%GPT-5.4Nearby rank
- 93.2%
DeepSeek V4.1 Flash 1017Nearby rank - 93.1%Kimi K3Similar price
- 90.7%GPT-5.5Predecessor
+2.7 pts vs GPT-5.5 (predecessor) · Ahead of 1 of 4 similarly priced models
OTIS Mock AIME 2024-2025
Best score per model · Higher is better
#13 of 158
- 100.0%Claude Sonnet 5.5Similar price
- 100.0%Qwen3.8 Max (0902)Similar price
- 99.7%Claude Fable 5.2Nearby rank
- 99.7%GPT-5.6 Terra
- 99.4%
Qwen3.8 MaxNearby rank - 99.4%DeepSeek V4.1 Flash 1017Nearby rank
- 97.2%Kimi K3Similar price
- 95.6%Gemini 3.1 Pro PreviewSimilar price
- 84.4%GPT-5.5Predecessor
+15.3 pts vs GPT-5.5 (predecessor) · Ahead of 2 of 4 similarly priced models
SimpleQA Verified
Best score per model · Higher is better
#46 of 87
- 73.5%Gemini 3.1 Pro PreviewSimilar price
- 63.0%GPT-5.5Predecessor
- 56.0%Grok 4.7Similar price
- 50.6%Kimi K3Similar price
- 46.5%Claude Sonnet 5.5Similar price
- 45.1%GPT-5.4Nearby rank
- 44.1%Qwen3.6 PlusNearby rank
- 43.2%GPT-5.6 Terra
- 42.9%DeepSeek V4.1 FlashNearby rank
- 42.5%DeepSeek V4.1 Flash 1017Nearby rank
-19.8 pts vs GPT-5.5 (predecessor) · Ahead of 0 of 4 similarly priced models
DeepSWE 1.1
Primary coding-agent benchmark · 113 tasks · Data generated 2026-09-22
- pass@4
- 88.5%
- Average cost / attempt
- $4.95
- Benchmark runs
- 4
Score vs cost by reasoning effort
Average cost per attempt (USD) against pass rate
All published configurations
| Harness / reasoning | pass@1 | pass@4 | Cost / attempt |
|---|---|---|---|
| mini-swe-agent · max | 69.6% | 88.5% | $4.95 |
| mini-swe-agent · xhigh | 60.2% | 80.5% | $2.13 |
| mini-swe-agent · high | 53.8% | 80.5% | $1.13 |
| mini-swe-agent · medium | 35.1% | 60.2% | $0.58 |
| mini-swe-agent · low | 24.1% | 44.2% | $0.43 |
pass@1 is the success rate across scored attempts; pass@4 is the share of tasks solved in at least one of four attempts. Costs are reported by DeepSWE for each configuration. These results are not directly comparable to evaluations using different harnesses. Official source JSON.
Reasoning effort
ARC-AGI-2
Score by reasoning effort · Higher is better
GPQA diamond
Score by reasoning effort · Higher is better
OTIS Mock AIME 2024-2025
Score by reasoning effort · Higher is better
All published scores
- Arena TextArena #511468
-13 vs GPT-5.5 (predecessor)
- xhigh146895% CI 1464 to 1472 · 48,825 votes
- Arena WebDevArena #611520
Level with GPT-5.5 (predecessor, overlapping 95% CIs)
- xhigh1520Codex Harness95% CI 1514 to 1526 · 15,482 votes
- Arena VisionArena #371266
-18 vs GPT-5.5 (predecessor)
- xhigh126695% CI 1259 to 1273 · 13,243 votes
- Arena DocumentArena #151474
Level with GPT-5.5 (predecessor, overlapping 95% CIs)
- xhigh147495% CI 1468 to 1481 · 11,412 votes
- Agent ArenaArena #35-0.018
- xhigh-0.01895% CI -0.027 to -0.010 · 3,844,264 observations
- ARC-AGI-2#16 of 8583.9%
-1.1 pts vs GPT-5.5 (predecessor)
- low18.8%
- medium37.5%
- high67.1%
- xhigh74.2%
- max83.9%
- FrontierMath-Tier-4-v2-Private#17 of 7270.7%
-1.8 pts vs GPT-5.5 (predecessor)
- max70.7%
Source:Epoch AI evaluation - FrontierMath-Tiers-1-3-v2-Private#13 of 8786.0%
+0.7 pts vs GPT-5.5 (predecessor)
- max86.0%
Source:Epoch AI evaluation - GPQA diamond#18 of 16193.3%
+2.7 pts vs GPT-5.5 (predecessor)
- none77.3%
- low87.4%
- max93.3%
Source:Epoch AI evaluation - OTIS Mock AIME 2024-2025#13 of 15899.7%
+15.3 pts vs GPT-5.5 (predecessor)
- none53.3%
- low88.9%
- max99.7%
Source:Epoch AI evaluation - SimpleQA Verified#46 of 8743.2%
-19.8 pts vs GPT-5.5 (predecessor)
- max43.2%
Source:Epoch AI evaluation
Benchmark compilation: Epoch AI, Capabilities & benchmarking (CC BY attribution). Scores are converted to percentages and model names matched to the Vertical Tabs catalog. Individual evaluations are credited to their original sources. External results may carry separate licences and evaluation configurations; do not interpret scores from different harnesses as directly comparable.
Arena ratings: Arena (lmarena-ai/leaderboard-dataset), licensed CC BY 4.0. Ratings come from pairwise human votes and are only comparable within one leaderboard; the interval is Arena's 95% confidence interval. Model names are matched to the Vertical Tabs catalog.