Frontier AI benchmark tracker: week of 2026-11-13
First week: no changes to compare yet. Every score comes from the linked publisher page. Scores from different benchmarks are not comparable with each other. Hover a row for details.
Coding
SWE Atlas (Overall) v1.1
| # | Model | Setup | Score | 95% interval | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| 1 |  Claude Fable 5.2 | 80.5% | 75.7% to 85.3% | leader | |||
| 2 | Claude Opus 5.5 | 77.4% | 72.4% to 82.4% | too close to call | |||
| 3 | Claude Fable 5.1 | 76.4% | 71.2% to 81.6% | too close to call | |||
| 4 |  Gemini 4 Carbon | 71.4% | 65.9% to 76.9% | too close to call | |||
| 5 | Gemini 4 Argon | 70.1% | 64.6% to 75.6% | behind | |||
| 6 |  GPT-6 Astra | 69.4% | 63.8% to 75.0% | behind | |||
| 7 | Claude Sonnet 5.5 | 68.3% | 62.7% to 73.9% | behind | |||
| 8 | GPT-6.1 Sol | 65.9% | 60.2% to 71.6% | behind | |||
| 9 |  Muse Wonder 1.0 | 64.8% | 59.0% to 70.6% | behind | |||
| 10 |  Qwen 3.8 | Max | 64.4% | 58.7% to 70.1% | behind | ||
| 11 |  DeepSeek V4.1 Flash | 62.1% | 56.3% to 67.9% | behind | |||
| 11 |  Grok 4.7 | 62.1% | 56.3% to 67.9% | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Scale says v1.1 scores are not comparable with v1.0.
Coding (issue fixing)
SWE-Bench Pro V2 (Full)
| # | Model | Setup | Score | 95% interval | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | Claude Code, xhigh | 99.4% | 99.0% to 99.8% | leader | ||
| 2 | Gemini 4 Argon | mini-swe-agent, high | 99.2% | 98.8% to 99.7% | too close to call | ||
| 2 | Claude Fable 5.1 | Claude Code, high | 99.1% | 98.6% to 99.6% | too close to call | ||
| 3 |  Kimi K3 | mini-swe-agent, max | 97.7% | 96.8% to 98.6% | behind | ||
| 4 | Inkling 2 Small | 97.7% | 96.2% to 99.1% | too close to call | |||
| 5 | GPT-6 Astra | Codex, high | 96.9% | 95.8% to 98.0% | behind | ||
| 5 | Muse Wonder 1.0 | mini-swe-agent, xhigh | 96.7% | 95.6% to 97.9% | behind | ||
| 6 | MiniMax M3.1 Flash | mini-swe-agent, xhigh | 96.6% | 95.4% to 97.7% | behind | ||
| 6 | GPT-6.1 Sol | mini-swe-agent, xhigh | 96.1% | 94.9% to 97.3% | behind | ||
| 7 | Claude Sonnet 5.5 | mini-swe-agent, xhigh | 95.8% | 94.5% to 97.1% | behind | ||
| 7 |  GLM 5.3 | mini-swe-agent, max | 95.6% | 94.3% to 96.9% | behind | ||
| 8 | GPT-5.6 Sol | Codex, xhigh | 95.5% | 94.1% to 96.9% | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Top score 99.4%; 3 other models are too close to call with #1.
Agentic coding
Terminal-Bench 4.0
| # | Model | Setup | Score | Change | Vs. #1 | |
|---|---|---|---|---|---|---|
| Kimi K3.1 Pro | 67.2% | leader | ||||
| Gemini 4 Carbon | High | 65.2% | no margin published | |||
| Claude Sonnet 5.5 | Max, Default Fallback | 63.6% | no margin published | |||
| Claude Fable 5.2 | Max, Default Fallback | 59.6% | no margin published | |||
| Claude Opus 5.5 | Max, Default Fallback | 59.6% | no margin published | |||
| GPT-6 Astra | Max | 59.1% | no margin published | |||
| Muse Wonder Full | Max | 57.1% | no margin published | |||
| GPT-6.1 Sol | Max | 56.1% | no margin published | |||
| GPT-6.1 Terra | Max | 55.6% | no margin published | |||
| Gemini 4 Hydrogen | Xhigh | 53.5% | no margin published | |||
| Claude Fable 5.1 | Max, Default Fallback | 52.0% | no margin published | |||
| Muse Wonder | Max | 52.0% | no margin published |
No margin published: dots only, so no tie can be judged.
Agentic tool use
MCP-Atlas
| # | Model | Setup | Score | 95% interval | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| 1 | Gemini 4 Argon | 90.7% | 88.9% to 92.5% | leader | |||
| 2 | Muse Spark 1.1 | 88.1% | 86.1% to 90.0% | too close to call | |||
| 3 | Claude Fable 5.1 | 87.2% | 85.2% to 89.2% | too close to call | |||
| 4 | Claude Opus 5 | xhigh | 85.8% | 83.7% to 87.9% | behind | ||
| 5 | Claude Opus 5.5 | 85.4% | 83.2% to 87.6% | behind | |||
| 6 | Inkling 2 Small | 85.4% | 83.2% to 87.6% | behind | |||
| 7 | Qwen3.8 2.4T A95B | xHigh | 84.5% | 82.2% to 86.8% | behind | ||
| 8 | GLM 5.3 | 84.2% | 82.0% to 86.4% | behind | |||
| 9 | Gemini 3.5 Flash | high | 83.6% | 81.3% to 85.9% | behind | ||
| 10 | Claude Fable 5 | 83.3% | 81.0% to 85.5% | behind | |||
| 11 | Kimi K3 | max | 82.3% | 80.0% to 84.6% | behind | ||
| 12 | Muse Spark | 82.2% | 79.9% to 84.5% | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Work
GDPval-AA v2.1
| # | Model | Setup | Score | 95% interval | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | Max, Default Fallback | 1866 | 1840 to 1892 | leader | |||
| Kimi K3.1 Pro | 1861 | 1833 to 1889 | too close to call | ||||
| Muse Wonder Full | Max | 1857 | 1832 to 1882 | too close to call | |||
| Claude Sonnet 5.5 | Max, Default Fallback | 1838 | 1814 to 1861 | too close to call | |||
| Claude Fable 5.2 | Max, Default Fallback | 1831 | 1804 to 1858 | too close to call | |||
| Muse Wonder | Max | 1770 | 1747 to 1793 | behind | |||
| Claude Fable 5.1 | Max, Default Fallback | 1763 | 1745 to 1782 | behind | |||
| DeepSeek V4.1 Flash 1017 | Max | 1746 | 1725 to 1767 | behind | |||
| Gemini 4 Carbon | High | 1733 | 1708 to 1757 | behind | |||
| Claude Opus 5 | Max | 1730 | 1711 to 1749 | behind | |||
| Grok 4.7 | Xhigh | 1717 | 1697 to 1737 | behind | |||
 Mimo V2.6 Pro | 1692 | 1672 to 1712 | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Work (frontier)
GDPx v1.0
| # | Model | Setup | Score | Cost (USD per task) | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| Claude Fable 5.2 | Claude Code, Max | 16.4% | $88.15 | leader | |||
| Claude Opus 5.5 | Claude Code, Max | 12.0% | $51.41 | no margin published | |||
| Gemini 4 Carbon | Antigravity, High | 10.9% | $16.29 | no margin published |
No margin published: dots only, so no tie can be judged.
Quoted from GDPx by SemiAnalysis (top 3 only); full leaderboard at gdpx.semianalysis.com.
GDPx by SemiAnalysis
Real-world agents
Agent Arena
| # | Model | Setup | Score | 95% interval | Cost (USD per task (median)) | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | Max | 0.303 | 0.264 to 0.343 | $7.43 | leader | ||
| 2 | Claude Opus 5.5 | Max | 0.195 | 0.175 to 0.215 | $5.47 | behind | ||
| 3 | Gemini 4 Carbon | High | 0.170 | 0.163 to 0.177 | $1.46 | behind | ||
| 4 | Claude Sonnet 5.5 | Max | 0.160 | 0.131 to 0.188 | $3.35 | behind | ||
| 5 | Claude Fable 5.1 | Max | 0.148 | 0.128 to 0.169 | $7.62 | behind | ||
| 6 | GPT-6 Astra | Max | 0.145 | 0.125 to 0.166 | $6.10 | behind | ||
| 7 | GPT-6.1 Sol | Max | 0.125 | 0.113 to 0.136 | $1.06 | behind | ||
| 8 | GPT-6.1 Terra | Max | 0.120 | 0.107 to 0.133 | $0.89 | behind | ||
| 9 | Gemini 4 Argon | High | 0.107 | 0.099 to 0.116 | $0.73 | behind | ||
| 10 | Muse Wonder Full | Max | 0.090 | 0.080 to 0.099 | behind | |||
| 11 | Claude Fable 5 | High | 0.073 | 0.061 to 0.085 | $3.34 | behind | ||
| 12 | Claude Opus 5 | Max | 0.072 | 0.059 to 0.085 | $4.34 | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Arena leaderboard dataset, CC-BY-4.0
Human preference (optional)
WebDev Arena
| # | Model | Setup | Score | 95% interval | Change | Vs. #1 | |
|---|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | xHigh | 1838 | 1817 to 1858 | leader | ||
| 2 | Claude Opus 5.5 | Max | 1815 | 1803 to 1826 | too close to call | ||
| 3 | Claude Sonnet 5.5 | xHigh | 1772 | 1762 to 1782 | behind | ||
| 4 | Gemini 4 Carbon | High | 1768 | 1753 to 1784 | behind | ||
| 5 | GPT-6 Astra | Max | 1764 | 1755 to 1772 | behind | ||
| 6 | GPT-6.1 Sol | Max | 1743 | 1734 to 1753 | behind | ||
| 7 | Claude Fable 5.1 | Max | 1742 | 1734 to 1750 | behind | ||
| 8 | Muse Wonder Full | Max | 1724 | 1709 to 1738 | behind | ||
| 9 | Claude Sonnet 5.5 | High | 1715 | 1707 to 1724 | behind | ||
| 10 | Gemini 4 Hydrogen | High | 1704 | 1690 to 1719 | behind | ||
| 11 | Claude Opus 5 | Max | 1691 | 1685 to 1697 | behind | ||
| 12 | Kimi K3.1 Pro | Max | 1687 | 1678 to 1697 | behind |
Dot = score, line = 95% interval, shaded band = the leader's interval, colour = the model's lab.
Arena leaderboard dataset, CC-BY-4.0
Method
- Trust labels. independent: a third party ran the models. crowd-preference: ranking from public human votes. Vendor-reported scores are never mixed in.
- Too close to call. Shown when a model's published 95% interval overlaps the leader's. Where a publisher shows no interval, the table says no margin published.
- GDPx score is GDPx's own formula (equal weight per domain and per partner) applied to the data GDPx publishes; it matches the numbers on gdpx.semianalysis.com. Credit: GDPx by SemiAnalysis.
- Arena data comes from Arena's official dataset (lmarena-ai/leaderboard-dataset, CC-BY-4.0).