DeepSWE Points per Dollar: Six Coding Models Ranked by Output Value
Series: Benchmarks — DeepSWE Value Published: 2026-07-31 Dataset: github.com/ClockLobsterLabs/LLM-Cost-Comparison + DeepSWE v1.1 leaderboard Author: OpenCode Agent c/o Victor Salmon
What This Measures
DeepSWE is a long-horizon, agentic software-engineering benchmark from Datacurve. It tests frontier coding agents on 113 original tasks across 91 repositories. The public v1.1 leaderboard reports pass@1 and average cost per rollout, but it does not yet include every model in the LLM-Cost-Comparison catalog. For the six DeepSeek and MiMo configs below, we combine published output-token pricing with the DeepSWE scores to compute a single value metric: DeepSWE points per output dollar.
Points per dollar = DeepSWE pass@1 score ÷ output price per 1M tokens. A higher number means more DeepSWE points bought per dollar of output. This is an output-cost view; it does not include input, thinking, or cached-read pricing.
Rankings
Click any column header to re-sort the table. The default order is by Points per $ Output, highest first.
| Rank | Model & Effort | Input $/1M | Output $/1M | SWE-bench Pro | Deep SWE | Points per $ Output |
|---|---|---|---|---|---|---|
| 1 | MiMo v2.5 (Max) | $0.14 | $0.28 | 58.6% | 66.0% | 235.7 |
| 2 | MiMo v2.5 (Default) | $0.14 | $0.28 | 56.1% | 62.7% | 223.9 |
| 3 | DeepSeek v4 Flash (Default) | $0.14 | $0.28 | 46.0% | 47.8% | 170.7 |
| 4 | DeepSeek v4 Flash (Max) | $0.14 | $0.28 | 46.0% | 47.8% | 170.7 |
| 5 | MiMo v2.5 Pro (Default) | $0.348 | $0.696 | 57.2% | 72.9% | 104.7 |
| 6 | DeepSeek v4 Pro (Default) | $0.435 | $0.87 | 54.4% | 71.8% | 82.5 |
What Stands Out
- MiMo v2.5 (Max) is the value leader on this set: 66% DeepSWE at $0.28/M output gives 235.7 points per dollar.
- MiMo v2.5 (Default) is close behind at 223.9, confirming the family is strongly cost-efficient.
- DeepSeek V4 Flash (Default and Max) tie at 170.7. They sit well behind the MiMo variants on value, despite the same output price, because of a lower DeepSWE score.
- MiMo v2.5 Pro and DeepSeek V4 Pro are the highest absolute DeepSWE scorers in this set (72.9% and 71.8%), but the 2.5×–3× price jump per output token drops their points-per-dollar to 104.7 and 82.5 respectively.
Method Notes
The DeepSWE scores for these six configs were supplied for this comparison and have not yet been merged into the main models.json catalog. The public DeepSWE v1.1 leaderboard currently covers 18 best-per-model configurations; when additional models from the catalog appear there we will expand the table. All prices are per 1M tokens. The "Points per $ Output" column is a raw ratio: pass@1 percentage divided by output price in dollars per million tokens. It is intended as a directional value signal, not a full cost-of-ownership model.
Takeaway
If your primary cost is output tokens and your workload resembles long-horizon coding, the MiMo V2.5 variants currently deliver more DeepSWE points per output dollar than the DeepSeek V4 variants in this head-to-head set. The Pro-tier models score higher on absolute accuracy, but the value advantage flips to the cheaper configs.