DeepSeek's latest model revision delivers low cost and strong benchmarks, but hallucination risks persist

V4.1 Flash revision 1017 scores near frontier labs on efficiency while struggling with factual reliability

By LineZotpaper

Published23 October 2026

Read Time3 min

DeepSeek has released revision 1017 of its V4.1 Flash open-weight model, achieving notable cost efficiency and task completion benchmarks that rival frontier models like GPT-5.6 Sol and Claude Opus 5, according to the company's internal data. However, independent testing reveals the model hallucinates 67% of the time when it does not know an answer, raising concerns about its suitability for tasks requiring factual accuracy.

DeepSeek's latest update to its V4.1 Flash model, revision 1017, promises significant improvements in efficiency and cost without sacrificing performance. The model is a Mixture-of-Experts design with 552 billion total parameters, 16 billion of which are active per run. It accepts text and image inputs and supports a context window of 1 million tokens, now standard for large language models.

According to DeepSeek's internal benchmarks, the revision slots between GPT-5.6 Sol and Claude Opus 5 in task tests, even outperforming Anthropic's model in some evaluations. The company highlights a dramatic reduction in hardware requirements: the V4.1 revision uses one-quarter of the high-bandwidth memory and one-eighth of the SSD storage for its KV cache compared to the previous V4 Flash, spending only 890 bytes per token entry.

Cost remains a major selling point. DeepSeek's API pricing for off-peak times is $0.15 per million input tokens and $0.60 per million output tokens. With cache hits, input costs drop to $0.003. The company also offers search functionality at $0.75 per 1,000 queries, undercutting competitors. Third-party providers may offer even lower rates.

Independent benchmarks from Artificial Analysis rank DSF-1017 as second in the large open-weight class with an intelligence score of 45, compared to Claude Opus 5.5's 58. The model is also among the fastest in its category, and its cost-per-task is edged out only by Xiaomi's MiMo V2.6-Pro.

Despite these strengths, the model shows significant reliability weaknesses. On Artificial Analysis's Omniscience benchmark, DSF-1017 answered correctly 43% of the time, competitive among open-weight models but well below frontier commercial models. Crucially, it hallucinated 67% of the answers it gave without factual support. The benchmark's overall index gave DSF-1017 a score of 5, compared to Kimi K3-max's 20 and frontier models' scores above 42.

DeepSeek's own marketing claims the model enables "wish-driven programming made real," suggesting suitability for open-ended tasks. However, the hallucination issue means careful task selection is advised. Tasks with verifiable outputs, such as programming, OCR with review, and format conversions, are safer because automated verification can catch errors. Open-ended research, fact-checking, and document drafting are better left to more reliable frontier models or humans.

The trade-off between cost and factuality is a growing theme among open-weight models. Analysts note that previous models like Kimi K3 offered similar value but also suffered from hallucination problems, suggesting a pattern where cost-efficiency and benchmark intelligence come at the expense of dependable factual recall.

Analysis

Why This Matters

DeepSeek's pricing undercuts frontier labs by orders of magnitude, potentially democratizing access to capable AI for small developers and cash-strapped organizations.
The hallucination problem undermines trust in AI for critical applications, and a pattern among open-weight models suggests this trade-off may be structural rather than solvable with current techniques.
If businesses and researchers rely on these models for factual tasks, the cost savings could be offset by the need for human oversight or verification systems.

Background

Large language models today fall broadly into two categories: expensive, highly reliable frontier models from companies like OpenAI and Anthropic, and cheaper, more accessible open-weight models from Chinese and other developers. DeepSeek has been a frontrunner in the open-weight space, releasing competitive models at a fraction of the cost. The V4.1 Flash revision continues this strategy, pushing performance up while driving hardware requirements down. However, reliability benchmarks have consistently shown that cost-optimised models struggle with knowing the limits of their knowledge, leading to confident but incorrect outputs. This tension between intelligence and trustworthiness is a central challenge for the field.

Key Perspectives

DeepSeek and open-weight advocates: The model offers frontier-competitive speed and cost-per-task, enabling agentic workflows and experimentation that would be prohibitively expensive with proprietary APIs. The hallucination issue is manageable with proper task design and verification. Developers and enterprise users: The low cost is attractive, but the risk of hallucination in open-ended or factual tasks limits practical use. Many will reserve DSF-1017 for verifiable tasks and rely on frontier models for high-stakes work, erasing some cost advantage. Critics and safety researchers: The 67% hallucination rate on non-answers is unacceptable for any application where accuracy matters. Without transparency or safeguards, deployment of such models could lead to real-world harm through misinformation or flawed decision-making.

What to Watch

Future updates from DeepSeek addressing hallucination, or the emergence of a third-party fine-tune that improves reliability.
Adoption by major AI agent platforms: if agents increasingly use DSF-1017, errors may cascade.
Pricing responses from frontier labs: will they cut costs to remain competitive, or will the market segment into reliability tiers?

Sources

Deepseek's latest AI model is smarter than ever, but its low running costs come at a price · Latest from Tom's Hardware

Checked

12 claims in this story checked against 1 source: 7 supported.

Show each claim

SupportedDeepSeek’s latest revision combines low running costs and strong benchmark results with hallucination risks.Latest from Tom's Hardware: “However, while the model is indeed cheap and scores well, it can be hallucination-prone.”
Partly supportedRevision 1017 scores near frontier labs on efficiency.Latest from Tom's Hardware: “even low-effort runs of DSF-1017 offer a task completion rate in the ballpark of “max” settings for pricey competitors.”
SupportedThe model hallucinates 67% of the time when it does not know an answer.Latest from Tom's Hardware: “That would be perfectly acceptable if it skipped answers without a foundation, but it hallucinated 67% of non-answers.”
Partly supportedThe revision improves efficiency and cost without sacrificing performance.Latest from Tom's Hardware: “The upgrade from V4 Flash to V4.1 Flash significantly slashed hardware requirements”
SupportedDeepSeek’s internal benchmarks place the model between GPT-5.6 Sol and Claude Opus 5, and show it beating Anthropic’s model in some task tests.Latest from Tom's Hardware: “According to Deepseek’s internal benchmarks, the model slots in neatly between GPT-5.6 Sol and Claude Opus 5, even beating Anthropic’s LLM in task tests.”
SupportedV4.1 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.Latest from Tom's Hardware: “V4.1 using only one-quarter of the HBM and one-eighth of SSD storage for its KV cache compared to its predecessor, spending only 890 bytes per token entry.”
SupportedOff-peak API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with cache-hit input costing $0.003.Latest from Tom's Hardware: “the price at off-peak times per million tokens is just $0.15 for input and $0.6 for output. On cache hits, the input token cost drops to a mere $0.003.”
SupportedDeepSeek search costs $0.75 per 1,000 queries and undercuts competitors.Latest from Tom's Hardware: “Deepseek also offers search functionality at $0.75 per 1,000 queries, an amount that undercuts competitors, with the closest one being Parallel at $1.”
SupportedArtificial Analysis ranks DSF-1017 second in the large open-weight class with a score of 45, compared with Claude Opus 5.5 at 58.Latest from Tom's Hardware: “Artificial Analysis ranks DSF-1017 (max) as only #2 on its intelligence index within the large open-weight class, with a score of 45. For reference, the most seemingly intelligent model, Claude Opus 5.5 (max with fallback), rings in at 58.”
Partly supportedOn Artificial Analysis’s Omniscience benchmark, DSF-1017 answered correctly 43% of the time, well below frontier commercial models.Latest from Tom's Hardware: “It answered questions correctly 43% of the time, which is among the best for open-weight models, but inferior to the proprietary competition.”
Partly supportedThe benchmark’s overall index scored DSF-1017 at 5, Kimi K3-max at 20, and frontier models above 42.Latest from Tom's Hardware: “DSF-1017 gets 5 points, compared to Kimi K3-max’s 20, with the frontier models coming in at 42+.”
Partly supportedAnalysts say earlier Kimi K3 offered similar value but also suffered from hallucination problems.Latest from Tom's Hardware: “My personal experience with Kimi was short-lived, as even short but precise fact look-ups came up with hallucinated information, and Artificial Analysis’ data seems to back this observation up.”