DeepSeek-V4.1-Flash-1017 hits original's max-effort score at default effort: $0.15 per correct task
ryanshrout · x · 2026-10-27
Signal65's PINNACLE benchmark tested the DeepSeek-V4.1-Flash-1017 October refresh: enterprises that passed on the original for needing max reasoning effort should re-evaluate.
Key findings:
- Efficiency: at default effort, 45% fewer weighted errors than the original at the same setting; matches the original's max-effort score (4,290) with 94% fewer output tokens per correct answer (20,829 vs 347,031) in 27 median rounds vs 43.5
- Vs July build: 33% fewer weighted errors than V4-Flash-0731; fabrication on 128K corpus down from 7.0% to 2.1%
- Concurrency: 312 concurrent agents at 10 tok/s each on one 8×B300 node — highest above 4,000 model score
- Cost: $0.15 per correct task, on the open-weight cost frontier; 93.9% enterprise jobs finished, 96.5% retrieval
- On the overall frontier, Claude Haiku 5.5 and GPT-6 Luna remain cheaper than any open-weight row
Methodology: live agentic sessions, deterministic code-verified scoring, no model judges, procedurally regenerated sandboxes per run.
More from Infra
- GPUs last 6 years, model weights go legacy in 6 months — bindureddy · 2026-10-27
- Block unveils Dessa, a local voice model capturing laughs, pauses and cries, trained on one NVIDIA GPU rack — technextpreneur · 2026-10-27
- Shipping a whole GPU rack to space? NVLink survival doubted — jwt0625 · 2026-10-27
- Surface Laptop Ultra review: impressive local AI but bandwidth limited — draginol · 2026-10-27
- Chamath: liquidated damages clauses will be the final boss for neocloud compute contracts — BrentM_BuildGuy · 2026-10-27
- AI demand is outpacing cloud supply, forcing major shifts in capacity reservations — DavidLinthicum · 2026-10-27