DeepSeek has released V4.1-Flash-1017, an updated Flash model for agentic and coding tasks that scores 77.0 percent on the DeepSWE benchmark at just under $0.20 per task. The model is available through the website, app and API, with weights published on Hugging Face and ModelScope. For business users this matters because agent runs that previously cost dollars now cost cents without loss of accuracy.
Benchmark gains and lower cost per run
Compared with V4.1-Flash, the new 1017 version posts higher benchmark scores, completes tasks faster and generates shorter outputs. DeepSeek describes the change as a Pareto improvement in the tradeoff between performance and cost. In its low setting the model still reaches more than 72 percent for about $0.10 per task, while the predecessor reached 74.2 percent at around $0.50 per run. Shorter outputs directly reduce API spend because fewer tokens are generated per completed task.
The model was tuned for development environments including DeepSeek Harness version v0.2.1-rc.1 or later, Tencent CodeBuddy and WorkBuddy. For end-to-end software creation, DeepSeek uses reinforcement learning with self-feedback, or RLSF. In this post-training method the model uses internal confidence scores or self-generated critiques as a reward signal, without human evaluators or external reward models. The company says this improves project development from game programming to 3D modeling and design, plus animation of compact explainer videos.
In company benchmarks, V4.1-Flash-1017 outperforms larger rivals such as Claude Opus 5, GPT 5.6-Sol, Kimi K3 and GLM 5.3 in several coding and agent categories. The cost chart for DeepSWE puts GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5 between $2 and more than $10 per task, with none matching the smaller DeepSeek model on accuracy. Independent benchmarks have not yet appeared, and benchmark scores alone do not predict developer satisfaction with code quality. That gap between lab results and production experience remains the key caveat.
What this means for automation budgets
For companies running coding agents at scale, the price shift changes which workflows pay off. At under $0.20 per DeepSWE task versus $2 to $10 for rival models, teams can afford more iterations, broader test coverage and parallel attempts within the same budget. Small firms gain access to agent-driven prototyping that was previously too expensive to run repeatedly. Large organizations can extend the same agent stack to more teams without a proportional increase in API costs.
Separately, DeepSeek opened its internal Search and Fetch APIs to outside developers for the first time. The company says the underlying engine already processes hundreds of millions of queries per day worldwide, with setup details in the API documentation. Each endpoint costs $0.75 per 1,000 calls, while DeepSeek API users get free access through server-side tools in DeepSeek Harness, Claude Code and Codex until October 31, 2026 at 4:00 PM UTC. Buyers should test retrieval quality, latency and data coverage in their own domain before committing.
A useful marker will be independent benchmark results and developer reports on code quality in real repositories over the next few months. A second marker is Chinese regulation of open-weight releases, reportedly tightening for systems ranked among the top 10 most capable models. If weights remain available and third-party tests confirm the cost advantage, low-cost agents will move faster into standard development pipelines.
