- Coding Upgrade: DeepSeek’s V4.1 Flash 1017 AI improves coding scores while still trailing Claude Opus 5.5 on factual reliability.
- Test Costs: Artificial Analysis estimates $0.26 per benchmark task, down from $0.40 for the predecessor at maximum reasoning effort.
- Cache Savings: The inherited September architecture compresses stored conversation state, reducing cache memory and storage requirements for long sessions.
- Web Tool Fees: Web search and page retrieval through DeepSeek’s supported integrations start charging October 31 at $0.75 per 1,000 successful requests each.
DeepSeek’s V4.1 Flash 1017 AI model improves coding performance and lowers average task costs in independent benchmarks, giving developers more capability for their spending. Evaluator Artificial Analysis found an uneven upgrade, however: its business-automation score slipped, and factual reliability still trailed premium rival Claude Opus 5.5.
Released on October 17, the update follows September’s V4.1 Flash release and is available on DeepSeek’s website, app and API. It accepts text and images, returns text, and can process up to 1 million tokens of context, the material supplied in a request. Its MIT-licensed model weights let developers download and adapt the trained model rather than rely solely on a hosted service.
Better Coding, Uneven Automation
Artificial Analysis’ comparison with the September model raises Flash’s Intelligence Index score from 39 to 45. That composite combines ten evaluations of tasks including coding, scientific reasoning and office work. Both DeepSeek versions were tested at maximum reasoning effort.
The practical coding gain is clearer in Terminal-Bench 4.0, where the score rises from 27% to 34%. These tasks require software to work through a terminal on jobs such as system administration and data processing. Artificial Analysis runs all 66 tasks three times using an agent framework called mini-swe-agent; a task passes only when every verification test passes.
Business workflows tell a different story. AutomationBench-AA, which asks models to operate simulated business applications through software interfaces, falls from 69% to 67%. Its score measures completed objectives, with a task receiving zero if the model breaks a guardrail.
DeepSeek’s own DeepSWE software-repair results point toward better performance with less spending. The company puts the update at 77% for just under $0.20 per task, versus 74.2% at about $0.50 for its predecessor. At low effort, it clears 72% for about $0.10. Those vendor results illustrate how reasoning settings and shorter outputs can change the price of an agent run.
DeepSeek says the update uses reinforcement learning with self-feedback to improve complete-project development, including games and 3D scenes.
🎮 Build complete projects from scratch
DeepSeek-V4.1-Flash-1017 uses reinforcement learning with self-feedback to improve end-to-end project development—from game creation to 3D modeling.
See it in action 👇
2/4 pic.twitter.com/Y5t4oGwq1x
— DeepSeek (@deepseek_ai) October 17, 2026
Factual Reliability Improves from a Low Base
On AA-Omniscience, which tests factual knowledge without web search, the update moves from minus 5 to 5. Anthropic’s Claude Opus 5.5 scores 46 in the tested configuration.
The benchmark asks 6,000 open-answer questions across 42 topics. Correct answers earn points, incorrect guesses lose points, and declining to answer is neutral. A score of zero means as many correct as incorrect answers; the scale runs from minus 100 to 100. Flash has moved onto the positive side, but remains much closer to that balance point than Opus.
That scoring exposes a different weakness from failing to complete a coding task. An attempted code change can be checked against tests. In factual recall, confidently supplying a wrong answer creates the error the benchmark penalizes.
Stronger Fact Scores at Different Prices
Xiaomi’s MiMo-V2.6-Pro, another downloadable model, costs half as much per weighted benchmark task while recording a similar terminal score and a higher factual-reliability score. DeepSeek leads it on business-workflow automation.
At higher prices, Opus 5.5 and Kimi K3.1 Pro deliver substantially stronger terminal and factual-reliability results. Kimi’s latest model, released October 23, is accessible through an API but has no public weights so far as of October 24
Artificial Analysis results as of October 23, 2026
| Model | Terminal-Bench 4.0 | AutomationBench-AA | AA-Omniscience | Cost per Task |
|---|---|---|---|---|
| V4.1 Flash, September | 27% | 69% | -5 | $0.40 |
| V4.1 Flash 1017 | 34% | 67% | 5 | $0.26 |
| MiMo-V2.6-Pro | 35% | 59% | 8 | $0.13 |
| Claude Opus 5.5 | 60% | 70% | 46 | $5.98 |
| Kimi K3.1 Pro | 67% | 73% | 30 | $2.11 |
DeepSeek models use max reasoning. The tested Opus configuration is Max, Default Fallback. Costs are weighted averages per Intelligence Index task; AA-Omniscience scores run from minus 100 to 100.
Artificial Analysis estimates Flash’s weighted task cost at about 35% below its predecessor’s, using the displayed $0.26 and $0.40 figures. Each evaluation’s token bill is divided by its task count, then weighted within the Intelligence Index. The estimate includes input, reasoning and answer tokens, plus cache pricing and typical cache-hit rates.
The update generated 211 million output tokens across the suite, against 253 million for September’s model. Both records use the same listed token rates. Lower total output therefore accompanies the cheaper evaluation.
Why Long Conversations Need Less Cache
The much-advertised hardware savings come from the September V4.1 architecture inherited by this revision. The comparison is with the V4-Flash model launched in April, an earlier generation. On the separate AA-LCR v1.1 long-document reasoning test, October’s revision scores 83%, nearly unchanged from September’s 84%.
DeepSeek’s technical report describes a smaller KV cache, the stored attention state that lets a model reuse earlier context. Sharing that state between layers and storing it at lower numerical precision cuts the global cache in high-bandwidth GPU memory to roughly one-quarter of V4-Flash’s footprint at equal sequence lengths.
Persistent cache storage on SSD or host memory falls to roughly one-eighth. Local attention state stays in a memory pool for only minutes during active sessions, while global state is retained for at least 72 hours. When local state expires, the system approximately rebuilds it by replaying a recent window of tokens, exchanging some computation for less stored state. DeepSeek says errors in choosing stored context or rebuilding local state could hurt performance in untested conditions.
What Changes the Developer’s Bill
The API rate card lists peak Flash prices of $0.30 per million uncached input tokens and $1.20 per million output tokens. Off-peak prices are $0.15 and $0.60. Cached input costs $0.006 per million tokens at peak and $0.003 off-peak. Tokens are the text units the service counts for billing, including generated reasoning.
Peak windows are 01:00–04:00 and 06:00–10:00 UTC on weekdays, excluding Chinese public holidays. Other hours, weekends and those holidays use the lower rates.
The inherited effort setting also changes how much reasoning the model produces. Thinking mode defaults to high, while the independent DeepSeek benchmark results cited here use max. In tool-using conversations, previous reasoning is passed back as context, adding material that the model processes on later turns.
DeepSeek’s Search and Fetch services give agents web access: Search returns page links and summaries, while Fetch retrieves a chosen page as text. Each service costs $0.75 per 1,000 successful requests; failed or empty requests are free.
As of October 23, these web-tool calls through DeepSeek Harness, Claude Code and Codex remain free for DeepSeek API users. Billing begins October 31 at 16:00 UTC.
