Fangcun DeepSeek 1017 Report - Shutdown Resistance cues - S07: leaderboard

Fangcun report 2610.35671v1, hosted API evaluations Oct 4–25 2026; Appendix A.1 exact model identities, maximum reasoning effort unless named otherwise, temperature 1/top_p 0.95. GLM 5.3 judge where required. API/format failures excluded by source; common cohort versus successfully completed denominators retained. Primary PDF printed values, not estimated plot heights. Figure 2.3 S1–6 standard task, S7–12 strong completion. Within each 6: question-tool/status-tool/user-warning pairs, odd script-path shown/even omitted. No explicit allow/prevent instruction; lower bypass source risk. Shared baseline: Figure 2.3/S07 and Figure 2.4/H01 report the same strong-completion, script-path-shown condition without an explicit allow/prevent instruction or an added remaining-work statement. Figure 2.3 identifies the question-tool warning; Figure 2.4 holds completion emphasis and script visibility fixed while varying the remaining-work statement. Its warning channel is inferred from that context and the identical seven-model vector. This board reports the documented baseline once and retains both panel references; individual trial reuse is not explicitly documented. Metric: Native rate (%).

Metric: Native rate (%). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.

Top models

#ModelScoreOverall rank
1GPT-6 Astra (Max)0#3 (GPT-6 Astra)
2DeepSeek V4.1 Flash (1017) (Max)44#25 (DeepSeek V4.1 Flash (1017))
3DeepSeek V4 Flash (Max)46#147 (DeepSeek V4 Flash)
4DeepSeek V4.1 Flash (Max)56#36 (DeepSeek V4.1 Flash)
5GPT-5.6 Sol (Max)66#19 (GPT-5.6 Sol)
6DeepSeek V4 Pro (Max)95#109 (DeepSeek V4 Pro)
7Gemini 3.8 Flash98.73#24

No result here: #2 Claude Opus 5.5, #6 GPT-6.1 Sol, #8 Gemini 4 Argon.

Interactive version: theaggregate.ai/benchmark?slug=fangcun-deepseek-1017-report-shutdown-resistance-cues-s07 · How It Works · Data refreshed daily, snapshot 2026-10-28.