DeepSeek V4.1 Flash 1017: $0.15/$0.6 per 1M tokens, 1M context, 67% hallucination rate on non-answers

DeepSeek has upgraded its V4.1 Flash open-weight model to revision 1017, pricing it at $0.15 per 1M input tokens and $0.6 per 1M output tokens at off-peak times. Tom's Hardware reports that the model scores well on intelligence and cost-per-task measures, but ranks among the lowest on Artificial Analysis' hallucination-sensitive Omniscience benchmark.

Specifications

• Architecture: Mixture-of-Experts (MoE), 552B total parameters, 16B active during runs
• Inputs: text and images (multimodal)
• Context window: 1 million tokens
• Weights: open-weight, so the model can be run on any infrastructure
• DeepSeek API pricing (off-peak): $0.15 per 1M input tokens, $0.6 per 1M output tokens, $0.003 per 1M input tokens on cache hits
• Search: $0.75 per 1,000 queries. The closest competitor is Parallel at $1.

The product page uses the slogan "wish-driven programming made real," which Tom's Hardware reads as a reference to vibe coding. Third-party inference providers have undercut DeepSeek's API in the past, so prices elsewhere could be lower.

Efficiency gains

According to DeepSeek, the upgrade from V4 Flash to V4.1 Flash cut hardware requirements significantly. V4.1 uses one-quarter of the HBM and one-eighth of the SSD storage for its KV cache compared with its predecessor, at 890 bytes per token entry.

DeepSeek's internal benchmarks, the company claims, place the model between GPT-5.6 Sol and Claude Opus 5, and ahead of Anthropic's model in task tests. DeepSeek also published a cost-per-task chart based on the DeepSWE dataset from September 2. Both the original DSF and DSF-1017 sit in the favorable top-left region. According to the company's chart, even low-effort runs of DSF-1017 reach task completion rates close to the "max" settings of pricier competitors. Tom's Hardware notes that DeepSWE is only one of many benchmarks.

Independent benchmarks

Artificial Analysis ranks DSF-1017 (max) second in the large open-weight class on its intelligence index, with a score of 45. Claude Opus 5.5 (max with fallback) leads with 58. The model is one of the fastest in its category, and only Xiaomi's MiMo V2.6-Pro beats it on cost-per-task.

The weakness is reliability. On Artificial Analysis' Omniscience benchmark, DSF-1017 answered questions correctly 43% of the time, among the best for open-weight models but below proprietary competitors. It also hallucinated on 67% of non-answers. The benchmark's main index rewards correct answers, penalizes wrong ones and ignores non-answers. DSF-1017 scores 5 on that index, against 20 for Kimi K3-max and 42 or higher for frontier models.

What this means

DSF-1017 combines a 1M token context window, low per-token pricing and a lower KV cache footprint than V4 Flash, according to DeepSeek. Those traits suit long-running agentic workloads. The Omniscience results show the cost: the model is poor at declining to answer when it lacks supporting facts.

Tom's Hardware concludes that tasks with verifiable outputs, such as programming, OCR with review and format conversion, are a good fit, because an agent can run verification steps and retry. Open-ended research, fact lookup and drafting of highly precise documents are better handled by a frontier model or a human, since hallucinations are unlikely to be caught. The author reports the same pattern with Kimi K3, where short fact lookups returned hallucinated information.

The author also argues that open-weight models remain important for making AI agents cheaper in absolute terms, and as a way to push down the cost of intelligence if frontier labs raise prices.