Benchmarks

Do more with less.

Structure is what makes a smaller model reason like a bigger one. On SWE-Atlas-QnA, the Coresource runtime running DeepSeek V4 Flash, an open model a fraction of the size, lands fifth of seventeen systems, matching agents that cost 14-32x more.

80.3%

Average pass@1 · #5 of 17 systems

$0.37

Cost per task · 14-32x cheaper

~$46

Full benchmark run · 124 tasks

Claude Fable 5

83.3

$11.75

Claude Opus 4.8

82.5

$7.70

Claude Opus 4.7

81.0

$5.64

Codex GPT-5.5

80.8

$5.07

Coresource · DeepSeek V4 Flash

80.3

$0.37

Codex GPT-5.5 med

79.1

$2.75

Opencode Opus 4.7

79.0

$2.93

Claude Opus 4.8 med

76.8

$3.26

Claude GLM-5.2

73.6

$6.47

Claude GLM-5.1

73.2

$4.33

Codex GPT-5.4 med

72.4

$2.27

Claude Opus 4.6 med

71.9

$1.26

Claude Opus 4.7 med

71.7

$1.68

Claude Sonnet 4.6

70.3

$0.27

Claude DeepSeek V4 Pro

67.8

$0.27

Claude Kimi K2.6

59.8

$1.18

Gemini 3.1 Pro

45.6

$2.00

Average pass@1 on SWE-Atlas-QnA (0-100); dollar figures are measured cost per task across 124 tasks. Source: Artificial Analysis.