Revision History

Update: added The Information's October 2026 report alleging that DeepSeek trained its latest model on GPT-6 Astra outputs, with a link to distillation attack

Update: added the 1017 checkpoint's Artificial Analysis Intelligence Index v5.0 score (32, October 29, 2026) and its FrontierCode 2.0 result

Auto-saved before update

Update: add Fangcun AI's October 26, 2026 third-party safety evaluation of DeepSeek-V4.1-Flash-1017 (methodology, selected results, recommendations, limitations) and a timeline row.

Added a scoped coding-session evaluation and methodological limitations

Added independent evaluation comparisons for the October checkpoint and self-hosting engine context

Update: Bash Recovery renamed Bash Reliability

Correction: Apex Bench v1.1 competitor values first published with 1017; change-log attribution; cost-chart sourcing (independent verification)

Update: add DeepSeek-V4.1-Flash-1017 (October 17, 2026) section with DeepSeek-reported benchmarks, harness and prompt-format notes, Search/Fetch API, timeline, and current API status; note NL2Repo value in the 1017 card

Improve internal links (2026-10-13)

Correction: AutomationBench-AA first place was as of September 11, 2026; by September 30 V4.1-Flash was third behind Claude Sonnet 5.5 and Opus 5.5

Added a short note on the DSec sandbox platform and the agent-sandbox/worker-container rollout split from the V4.1-Flash technical report (Sections 5.1.2-5.1.3), with a link to the new DSec page.

Auto-saved before update

Auto-saved before update

No summary provided