Unverified report: DeepSeek V4 Pro quietly updated, hard agent task pass rate jumps 16 points
teortaxesTex · x · 2026-11-08
The most credible report yet (unverified, not official) of a quiet DeepSeek V4 Pro update. After a service interruption, LobeHub's Max re-ran his private LobeBench Hard 50 eval (50 hard agent tasks from real scenarios, 3 attempts each):
- Current V4 Pro: 64% (32/50), ranked #1 among Chinese models, up from 48% (24/50, #13) for the 0813 version — a 16-point jump.
- Counting at-least-one-success-in-three, the two versions score 82% vs 80% — nearly identical. The gap is in completing tasks on all three consecutive tries, i.e. execution stability on complex agent tasks.
He cautions that the full leaderboard is needed and the eval shape is suspect — do all models fail the same 9-10 tasks?
Related event: DeepSeek V4 Pro疑似更新引发能力提升争议(2 posts)→
More from Models
- Grok Ultra mocked as "food ration" tokens: same price lets Anthropic run a small civilization — mertdumenci · 2026-11-08
- 1,400 Claude Code sessions suggest Opus 5.5's reasoning shifted abruptly on Nov 3 — churchkey · 2026-11-08
- Leaked Qwen4-27B early build: ~35GB, tops benchmark, may be sparse MoE — teortaxesTex · 2026-11-08
- Muse token pricing: 2B tokens for $35, plus 1B-token referral code — RichardsonDx · 2026-11-08
- Same model, same config: Reddit user documents rare episodes where agents act like a lousy support rep — No_Dragonfruit_1599 · 2026-11-08
- User frustrated as Anthropic monthly promo credits expire within 30 days — arthurcolle · 2026-11-08