Comparison
Click any column to sort. Default: effective cost per score point, lowest first.
| Model ↕ | Effort / Harness ↕ | Overall coding score ↕ | DeepSWE ↕ | TerminalBench ↕ | SWE Atlas ↕ | API / PAYG $/task ↕ | Credits at -20% ↕ | Codex Pro 20x est. ↕ | Effective ¢ / score point ↕ |
|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4 Flash 0731 | Max / Codex | 50 | 43% | 68% | 39% | $0.06 | - | - | 0.12¢ |
| Luna | Medium / Codex | 42 | 37% | 62% | 27% | $0.09 | $0.07 | $0.05-$0.06 | 0.13¢ |
| Luna | High / Codex | 52 | 53% | 73% | 29% | $0.18 | $0.14 | $0.09-$0.11 | 0.19¢ |
| DeepSeek V4 Pro 0813 | Max / Codex | 43 | 16% | 63% | 49% | $0.09 | - | - | 0.21¢ |
| Composer 2.5 | Standard / Cursor CLI | 38 | 16% | 68% | 31% | $0.09 | - | - | 0.24¢ |
| Luna | XHigh / Codex | 53 | 57% | 71% | 31% | $0.24 | $0.19 | $0.12-$0.15 | 0.25¢ |
| Luna | Max / Codex | 57 | 63% | 75% | 33% | $0.29 | $0.23 | $0.15-$0.18 | 0.29¢ |
| Terra | Medium / Codex | 48 | 46% | 70% | 29% | $0.67 | $0.54 | $0.34-$0.42 | 0.79¢ |
| Terra | High / Codex | 55 | 60% | 72% | 31% | $1.14 | $0.91 | $0.57-$0.71 | 1.16¢ |
| Terra | XHigh / Codex | 56 | 58% | 77% | 33% | $1.36 | $1.09 | $0.68-$0.85 | 1.37¢ |
| Composer 2.5 | Fast / Cursor CLI | 38 | 16% | 68% | 31% | $0.56 | - | - | 1.47¢ |
| Terra | Max / Codex | 60 | 67% | 78% | 36% | $1.93 | $1.54 | $0.97-$1.21 | 1.82¢ |
| Sol | Medium / Codex | 62 | 64% | 81% | 40% | $2.19 | $1.75 | $1.10-$1.37 | 1.99¢ |
| Sol | High / Codex | 64 | 65% | 82% | 45% | $3.00 | $2.40 | $1.50-$1.88 | 2.64¢ |
| Sol | XHigh / Codex | 63 | 67% | 80% | 43% | $3.74 | $2.99 | $1.87-$2.34 | 3.34¢ |
| Grok 4.5 | High / Grok Build | 64 | 60% | 84% | 48% | $2.44 | - | - | 3.81¢ |
| Claude Opus 5 | Low / Claude Code | 59 | 57% | 82% | 39% | $2.30 | - | - | 3.90¢ |
| Astra | Max / Codex | 67 | 67% | 83% | 51% | $4.72 | $3.78 | $2.36-$2.95 | 3.96¢ |
| Sol | Max / Codex | 65 | 69% | 83% | 43% | $5.00 | $4.00 | $2.50-$3.13 | 4.33¢ |
| Kimi K3 | Kimi Code CLI | 63 | 64% | 88% | 37% | $3.08 | - | - | 4.89¢ |
| Claude Opus 5 | Medium / Claude Code | 64 | 63% | 85% | 44% | $3.17 | - | - | 4.95¢ |
| Claude Opus 5 | High / Claude Code | 66 | 61% | 87% | 49% | $3.92 | - | - | 5.94¢ |
| Claude Opus 5 | XHigh / Claude Code | 68 | 60% | 89% | 55% | $8.17 | - | - | 12.01¢ |
| Claude Fable 5.1 | Max + fallback / Claude Code | 70 | 66% | 89% | 56% | $9.18 | - | - | 13.11¢ |
| Claude Opus 5 | Max / Claude Code | 67 | 63% | 89% | 49% | $8.94 | - | - | 13.34¢ |
Notes
Snapshot: . Scores and measured API cost come from Artificial Analysis Coding Agent Index v1.4.
Codex estimates: Credits applies the working 20% discount. Pro 20x assumes included compute delivers about 1.6x to 2.0x the value of base pay-as-you-go credits. Neither is an official per-task quote.
Effective cost per score point: effective task cost divided by Overall coding score, shown in cents. Codex models use the midpoint of the Pro 20x estimate; other models use API / PAYG cost. It is a cost-efficiency indicator, not a linear measure of intelligence.
DeepSeek on OpenRouter: Flash 0731 currently starts near $0.05/M input and $0.10/M output, versus DeepSeek Direct at $0.22-$0.44/M fresh input and $0.66-$1.32/M output. The table keeps the measured $0.06 benchmark task cost because no identical token trace has been rerun through OpenRouter. See current OpenRouter endpoints.
Missing current models: Grok 4.6 and Claude Sonnet 5 have CursorBench results, but no comparable Coding Agent Index v1.4 run yet. Mixing those scores into this ranking would make the decision less reliable.
Related guide: How to Have an Endless Supply of Tokens in Codex.