LetTokenBurnSWE-Duel: Towards A Self-Scaling Arena for Coding Agents
A self-scaling arena where coding agents generate and repair bugs in real repositories; evaluated via 1,500 turns across 11 participants on 15 repos.
$ npm i -g lettokenburn-cli$ ltb bench pull swe-duel
External harness. Requires the SWE-bench style execution harness; run with the official harness or an agent scaffold. https://github.com/SWE-bench/SWE-bench
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| GPT-5.5 | 2.8 | codex | Independent | medium | Oct 8, 2026 | arxiv.org |
| Claude Opus 4.8 | 2.5 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
| GPT-5.5 | 2.1 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
| Kimi K2.7 Code | 1.8 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
| DeepSeek V4 Pro | 1.5 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
| GLM-5.2 | 1.2 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
| Qwen3.7 Max | 0.5 | mini-swe-agent | Independent | medium | Oct 8, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Agents engage in mirrored Red-Blue turns: Red implements features with embedded bugs, Blue repairs them. Tasks are created at evaluation time to prevent leakage.
Scoring relies on win rates in adversarial rounds. GPT-5.5 ranked highest as solver (98.3% repair), while Claude Opus 4.8 led as task author (27.3% Red win rate).