Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs

Deepseek released Deepseek-V4.1-Flash-1017 on its website, app and API, and published the model weights openly on Hugging Face and ModelScope. The update replaces V4.1-Flash and, according to the company, improves on its predecessor in every tested category: higher benchmark scores, faster task completion and shorter outputs. Deepseek describes the release as a Pareto improvement in the performance-versus-cost tradeoff, meaning it claims gains on both axes rather than trading one for the other.
The post-training method behind the model is reinforcement learning with self-feedback, or RLSF. In this setup the model uses its own internal confidence scores or self-generated critiques as the reward signal instead of relying on human evaluators or a separate external reward model. This is a notable design choice for developers: it lowers the cost of iterating on post-training, since the feedback loop is generated by the model itself. The method is aimed at end-to-end project development, from game programming to 3D modeling and design work. Deepseek also says the model can animate compact explainer videos.
Deepseek says the model was tuned for development environments including its own Deepseek Harness, where version v0.2.1-rc.1 or later is recommended, along with Tencent CodeBuddy and WorkBuddy.
The cost numbers are the most concrete part of the release. On DeepSWE, Deepseek's chart shows V4.1-Flash-1017 reaching a top score of 77.0 percent at an average cost just under $0.20 per task. In its low setting it still clears 72 percent for roughly $0.10 per run. The predecessor, V4.1-Flash, scored 74.2 percent at around $0.50 per run. The same chart places rivals including GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 between $2 and more than $10 per task, and none of them match the smaller Deepseek model's accuracy in that comparison. Deepseek also says the model outperforms larger competitors such as Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 across several coding and agent categories.
All of these figures come from Deepseek's own benchmarks, and independent results have not yet been published. This is the key caveat for anyone planning around them. Benchmark spreads of a few points between models often shrink or reverse under external evaluation, and per-task cost depends heavily on the harness, the task mix and output length. The short-output design is clearly central to the cost story: fewer generated tokens translate directly into lower per-task spend, but shorter outputs can also mean terser reasoning traces, which is exactly the kind of thing that shows up in real developer experience rather than in scores. How the model handles code quality and edge cases in practice is a separate question from benchmark accuracy, and developers often form very different impressions than leaderboards suggest.
Separately, Deepseek is making its in-house web search engine available to outside developers through new Search and Fetch APIs. The company says the engine already handles hundreds of millions of queries per day worldwide and was built internally. Both endpoints are priced at $0.75 per 1,000 calls, and setup details are in the API documentation. Users of the Deepseek API get free access to these tools when they run through server-side tooling in Deepseek Harness, Claude Code and Codex, a promotion that lasts until October 31, 2026 at 4:00 PM UTC. For teams building agent workflows, a cheap, high-volume search-and-fetch pair is a meaningful building block: retrieval is one of the recurring costs in agentic systems, and having it bundled with the model provider can simplify architecture. The free window through third-party coding agents like Claude Code and Codex is also an unusual move, since it puts Deepseek's search inside competitors' tools rather than requiring users to adopt Deepseek's own harness.
There is a wider context worth flagging. Reports indicate the Chinese government plans to tighten rules on publishing open-weight models, and may even block releases outright, with one deciding factor being whether a system ranks among the top 10 most capable models. If enacted, that would affect the open-weight distribution channel that Deepseek has relied on. This is a reported regulatory plan, not a confirmed restriction on this release, and it is unclear how it would apply in practice. A likely trade-off for teams is that open-weight availability, which makes a model attractive for self-hosting and cost control, could become less predictable for Chinese labs over time, pushing some organizations toward hosted API access or fallback models.
Why it matters: Developers building coding agents and end-to-end software tools now have an open-weight option whose claimed cost per task, roughly $0.10 to $0.20 on Deepseek's own chart, sits far below rivals priced at $2 to over $10, which could materially change unit economics for high-volume agent workloads. The new Search and Fetch APIs at $0.75 per 1,000 calls, with free access through Deepseek Harness, Claude Code and Codex until October 31, 2026, give teams a cheap retrieval layer without leaving their existing tools. The main caution is inferential: until independent benchmarks and real-world code-quality feedback arrive, teams should validate the model on their own task mix before committing production pipelines to it.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Deepseek has released V4.1-Flash-1017 as an open-weight model that posts better benchmarks, finishes tasks faster, and produces shorter outputs than its predecessor. The company is also opening up its internal search engine through new Search and Fetch APIs at $0.75 per 1,000 calls. API users in tools like Claude Code and Codex get free access for now. The article Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs appeared first on The Decoder .