DeepSeek lowers running costs, but hallucinations remain a concern

DeepSeek has upgraded its V4.1 Flash open-weight model to revision 1017, pairing strong task performance with low inference costs. The model accepts text and images and supports a context window of 1 million tokens. Compared with V4 Flash, V4.1 Flash uses only one-quarter of the HBM and one-eighth of SSD storage for its KV cache, requiring 890 bytes per token entry.

Off-peak pricing per million tokens is $0.15 for input and $0.6 for output, with cached input costing $0.003. DeepSeek’s internal benchmarks place the model between GPT-5.6 Sol and Claude Opus 5. Independent testing from Artificial Analysis gives DSF-1017 (max) an intelligence index score of 45, compared with 58 for Claude Opus 5.5 (max with fallback). Its cost-per-task trails only Xiaomi’s MiMo V2.6-Pro.

Reliability remains a significant limitation. In Artificial Analysis’ Omniscience benchmark, the model answered correctly 43% of the time but hallucinated 67% of non-answers, indicating difficulty withholding unsupported responses. That makes programming, reviewed OCR and format conversions more suitable uses because outputs can be checked and corrected. Open-ended research, factual lookups and highly precise documents carry greater risks, despite the model’s attractive economics.