Home / Benchmarks / Coding Index
Lumina Coding Index By Lumina Index designed by LuminaBench using attributed, publisher-run benchmarks. Lumina did not rerun those evaluations. CursorBench and Terminal-Bench, combined.
Experimental Verified 30 October 2026
CursorBench 4.0 50% weight Terminal-Bench 4.0 50% weight Scores marked Est. include an estimated result. Compare the sources below.
Lumina Coding Index CursorBench 50% · Terminal-Bench 50%.
Observed and estimated results · 20 of 65 configurations
0 300 600 900 1,200 Claude Opus 5.5 Max · Cross-benchmark evidence · 1135 1,135 Claude Opus 5.5 Max · 2/2 Claude Sonnet 5.5 Max · Cross-benchmark evidence · 1118 1,118 Claude Sonnet 5.5 Max · 2/2 Claude Opus 5.5 xHigh · Cross-benchmark evidence · 1115 1,115 Claude Opus 5.5 xHigh · 1/2 Est. Claude Opus 5.5 High · Cross-benchmark evidence · 1115 1,115 Claude Opus 5.5 High · 1/2 Est. GPT-6.1 Sol Max · Cross-benchmark evidence · 1110 1,110 GPT-6.1 Sol Max · 1/2 Est. GPT-6 Astra Max · Cross-benchmark evidence · 1110 1,110 GPT-6 Astra Max · 1/2 Est. GPT-6 Astra xHigh · Cross-benchmark evidence · 1109 1,109 GPT-6 Astra xHigh · 1/2 Est. GPT-6 Astra High · Cross-benchmark evidence · 1109 1,109 GPT-6 Astra High · 1/2 Est. Claude Fable 5.1 Max · Cross-benchmark evidence · 1094 1,094 Claude Fable 5.1 Max · 2/2 Claude Fable 5.1 xHigh · Cross-benchmark evidence · 1093 1,093 Claude Fable 5.1 xHigh · 2/2 GPT-6 Astra Medium · Cross-benchmark evidence · 1093 1,093 GPT-6 Astra Medium · 1/2 Est. Claude Sonnet 5.5 xHigh · Cross-benchmark evidence · 1090 1,090 Claude Sonnet 5.5 xHigh · 1/2 Est. Gemini 4 Argon High · Cross-benchmark evidence · 1088 1,088 Gemini 4 Argon High · 1/2 Est. Gemini 4 Argon Medium · Cross-benchmark evidence · 1087 1,087 Gemini 4 Argon Medium · 1/2 Est. Claude Opus 5.5 Medium · Cross-benchmark evidence · 1085 1,085 Claude Opus 5.5 Medium · 1/2 Est. GPT-6 Astra Low · Cross-benchmark evidence · 1077 1,077 GPT-6 Astra Low · 1/2 Est. Claude Fable 5.1 High · Cross-benchmark evidence · 1076 1,076 Claude Fable 5.1 High · 2/2 GPT-6 Sol Max · Cross-benchmark evidence · 1071 1,071 GPT-6 Sol Max · 1/2 Est. Gemini 4 Argon Low · Cross-benchmark evidence · 1069 1,069 Gemini 4 Argon Low · 1/2 Est. Muse Wonder Max · Cross-benchmark evidence · 1067 1,067 Muse Wonder Max · 1/2 Est.
■ Claude Fable 5.1■ Claude Opus 5.5■ Claude Sonnet 5.5■ Gemini 4 Argon■ GPT-6 Astra■ GPT-6 Sol■ GPT-6.1 Sol■ Muse Wonder
Lumina Coding Index 20 configurations · Lumina Coding Index 2 · coding-index-v2 · Experimental LuminaBench 0 300 600 900 1,200 Claude Opus 5.5 Max · Cross-benchmark evidence · 1135 1,135 Claude Opus 5.5 Max · 2/2 Claude Sonnet 5.5 Max · Cross-benchmark evidence · 1118 1,118 Claude Sonnet 5.5 Max · 2/2 Claude Opus 5.5 xHigh · Cross-benchmark evidence · 1115 1,115 Claude Opus 5.5 xHigh · 1/2 Est. Claude Opus 5.5 High · Cross-benchmark evidence · 1115 1,115 Claude Opus 5.5 High · 1/2 Est. GPT-6.1 Sol Max · Cross-benchmark evidence · 1110 1,110 GPT-6.1 Sol Max · 1/2 Est. GPT-6 Astra Max · Cross-benchmark evidence · 1110 1,110 GPT-6 Astra Max · 1/2 Est. GPT-6 Astra xHigh · Cross-benchmark evidence · 1109 1,109 GPT-6 Astra xHigh · 1/2 Est. GPT-6 Astra High · Cross-benchmark evidence · 1109 1,109 GPT-6 Astra High · 1/2 Est. Claude Fable 5.1 Max · Cross-benchmark evidence · 1094 1,094 Claude Fable 5.1 Max · 2/2 Claude Fable 5.1 xHigh · Cross-benchmark evidence · 1093 1,093 Claude Fable 5.1 xHigh · 2/2 GPT-6 Astra Medium · Cross-benchmark evidence · 1093 1,093 GPT-6 Astra Medium · 1/2 Est. Claude Sonnet 5.5 xHigh · Cross-benchmark evidence · 1090 1,090 Claude Sonnet 5.5 xHigh · 1/2 Est. Gemini 4 Argon High · Cross-benchmark evidence · 1088 1,088 Gemini 4 Argon High · 1/2 Est. Gemini 4 Argon Medium · Cross-benchmark evidence · 1087 1,087 Gemini 4 Argon Medium · 1/2 Est. Claude Opus 5.5 Medium · Cross-benchmark evidence · 1085 1,085 Claude Opus 5.5 Medium · 1/2 Est. GPT-6 Astra Low · Cross-benchmark evidence · 1077 1,077 GPT-6 Astra Low · 1/2 Est. Claude Fable 5.1 High · Cross-benchmark evidence · 1076 1,076 Claude Fable 5.1 High · 2/2 GPT-6 Sol Max · Cross-benchmark evidence · 1071 1,071 GPT-6 Sol Max · 1/2 Est. Gemini 4 Argon Low · Cross-benchmark evidence · 1069 1,069 Gemini 4 Argon Low · 1/2 Est. Muse Wonder Max · Cross-benchmark evidence · 1067 1,067 Muse Wonder Max · 1/2 Est.
Lumina Score vs Cost per Task Published API-equivalent USD per task.
8 of 19 families
■ Claude Fable 5.1▼ Claude Haiku 5.5■ Claude Opus 5■ Claude Opus 5.5⬢ Claude Sonnet 5⬢ Claude Sonnet 5.5● Composer 2.5⬢ DeepSeek V4.1 Flash 1017✚ Gemini 3.8 Flash▲ Gemini 4 Argon◆ GLM 5.3● GLM 5.3 Flash■ GPT-5.6 Luna● GPT-5.6 Sol■ GPT-5.6 Terra▲ Grok 4.7▲ Grok 4.6◆ Muse Spark 1.3● Muse Wonder
LuminaBench Lumina Score vs Cost per Task Published API-equivalent USD per task. · CursorBench 4.0 · cursorbench-4-z75-v1 · Experimental $0.1 $0.3 $1 $3 $10 800 900 1,000 1,100 1,200 Average API-equivalent cost/task · USD (log scale) Lumina Score Claude Opus 5.5 Max · CursorBench publisher-run agent evaluation · 1161 Claude Opus 5.5 xHigh · CursorBench publisher-run agent evaluation · 1146 Claude Opus 5.5 High · CursorBench publisher-run agent evaluation · 1146 Claude Sonnet 5.5 Max · CursorBench publisher-run agent evaluation · 1141 Claude Sonnet 5.5 xHigh · CursorBench publisher-run agent evaluation · 1121 Gemini 4 Argon High · CursorBench publisher-run agent evaluation · 1119 Gemini 4 Argon Medium · CursorBench publisher-run agent evaluation · 1118 Claude Opus 5.5 Medium · CursorBench publisher-run agent evaluation · 1116 Claude Fable 5.1 Max · CursorBench publisher-run agent evaluation · 1110 Claude Fable 5.1 xHigh · CursorBench publisher-run agent evaluation · 1108 Gemini 4 Argon Low · CursorBench publisher-run agent evaluation · 1100 Muse Wonder Max · CursorBench publisher-run agent evaluation · 1098 Claude Fable 5.1 High · CursorBench publisher-run agent evaluation · 1088 Claude Haiku 5.5 Max · CursorBench publisher-run agent evaluation · 1081 Muse Wonder xHigh · CursorBench publisher-run agent evaluation · 1079 Claude Sonnet 5.5 High · CursorBench publisher-run agent evaluation · 1076 DeepSeek V4.1 Flash 1017 Max · CursorBench publisher-run agent evaluation · 1075 Claude Fable 5.1 Medium · CursorBench publisher-run agent evaluation · 1067 Claude Opus 5 Max · CursorBench publisher-run agent evaluation · 1065 DeepSeek V4.1 Flash 1017 Medium · CursorBench publisher-run agent evaluation · 1063 Claude Opus 5 xHigh · CursorBench publisher-run agent evaluation · 1061 Muse Wonder High · CursorBench publisher-run agent evaluation · 1061 DeepSeek V4.1 Flash 1017 High · CursorBench publisher-run agent evaluation · 1057 Muse Wonder Medium · CursorBench publisher-run agent evaluation · 1053 Claude Fable 5.1 Low · CursorBench publisher-run agent evaluation · 1053 DeepSeek V4.1 Flash 1017 Low · CursorBench publisher-run agent evaluation · 1051 Claude Opus 5 High · CursorBench publisher-run agent evaluation · 1049 Claude Haiku 5.5 xHigh · CursorBench publisher-run agent evaluation · 1046 Claude Opus 5.5 Low · CursorBench publisher-run agent evaluation · 1041 Claude Opus 5 Medium · CursorBench publisher-run agent evaluation · 1037 Claude Haiku 5.5 High · CursorBench publisher-run agent evaluation · 1029 Muse Wonder Low · CursorBench publisher-run agent evaluation · 1023 Claude Opus 5 Low · CursorBench publisher-run agent evaluation · 1015 Claude Sonnet 5.5 Medium · CursorBench publisher-run agent evaluation · 1002 Claude Haiku 5.5 Medium · CursorBench publisher-run agent evaluation · 982 Claude Sonnet 5.5 Low · CursorBench publisher-run agent evaluation · 973 Muse Wonder Minimal · CursorBench publisher-run agent evaluation · 959 Claude Haiku 5.5 Low · CursorBench publisher-run agent evaluation · 931 Claude Opus 5.5 Claude Sonnet 5.5 Gemini 4 Argon Claude Fable 5.1 Muse Wonder Claude Haiku 5.5 DeepSeek V4.1 Flash 1017 Claude Opus 5 Claude Fable 5.1 Claude Haiku 5.5 Claude Opus 5 Claude Opus 5.5 Claude Sonnet 5.5 DeepSeek V4.1 Flash 1017 Gemini 4 Argon Muse Wonder Method & coverage Complete scores are a 50/50 average of CursorBench and Terminal-Bench. Each source keeps its own published score, harness and resource figures. This index is separate from the Capabilities Index.
Est. marks a score with one missing benchmark. A fixed adjustment of 30.7 points, based on 20 paired configurations across 12 model families, puts those estimates on the combined scale. Hover to see the measured and estimated contributions.
Leaving each family out in turn reduced average backtest error from 30.7 to 15.3 points. The ±33 band is a retrospective error range, not a confidence interval or a guarantee for unseen models.
20 configurations have both results; 45 have supported estimates. The remaining 31 have evidence only on their individual benchmark pages because their scores fall outside the reference range or their identities need review. Only complete results receive qualifying ranks.
Method: coding-index-v2.