DeepSeek Wants Better AI at Lower Cost

Editor’s note: This article was written by Lucia, a TechNode reporter.

Building a more capable AI model is only half the battle. The other half is making it affordable to run at scale. DeepSeek announced a V4.1 Flash update aimed at faster task completion and shorter outputs, alongside an open-source framework for more efficient inference.

A Model Update With Promising Early Results

On October 17, DeepSeek released V4.1-Flash-1017, an update to its V4.1 Flash model. The company's published benchmark results show improvements in several areas, including coding and agentic tasks. For example, its Terminal-Bench 4.0 score rose from 31.2 to 42.3, while DeepSWE v1.1 increased from 74.2 to 77.0.

A cost-performance chart for the DeepSWE v1.1 coding benchmark shows V4.1 Flash-1017 achieving a 77% Pass@1 score at about $0.20 per task, outperforming several models that cost substantially more to run. At lower-cost settings of $0.11–$0.14 per task, it still scores around 72–73%.

DeepSeek also reports shorter task execution times and outputs, which can reduce inference costs when a model completes the same work using fewer tokens and less computing time. The update is optimized for agent frameworks including DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.

These figures, however, are primarily the company's own benchmark results, and different tests use different frameworks and evaluation settings. They do not establish that the model will perform equally well across all applications.

A separate signal came from Arena's Agent Arena leaderboard. According to ITHome's October 20 report, the updated model ranked first among open-weight models and 11th overall, based on 7,498 real-use sessions. Its median cost was approximately $0.18 per task, although different leaderboard pages showed figures ranging from $0.18 to $0.19.

Opening Up the Infrastructure

On October 20, DeepSeek released DeepInfer, its internal inference framework, alongside deepseek-recipe, a library for processing API requests, preparing model prompts and parsing outputs.

Inference frameworks manage how models use computing resources to handle requests. Their design affects latency, throughput and the hardware required to serve users, making them an important — if less visible — factor in AI operating costs.

DeepSeek describes DeepInfer as the framework powering its large-scale production services. The open-source release includes optimized computing kernels, asynchronous scheduling and support for distributed inference and speculative decoding, according to its README and deployment documentation. It currently supports Nvidia SM100 GPUs and Huawei Ascend 950 NPUs, with DeepSeek V4.1 Flash listed as its supported model.

The release is not a complete inference stack. DeepSeek’s documentation identifies a KV-cache system, prefill/decode disaggregation, and load balancing across data-parallel replicas as requirements for a complete service that are not fully covered by the release, although most of the core disaggregation functionality is already implemented.

The open-source code gives developers access to parts of DeepSeek’s model-serving technology. The practical benefits will depend on their hardware, workloads, and the additional engineering needed to deploy it.

Lower Costs Are Not the Whole Story

At its September launch of V4.1 Flash, DeepSeek reported reducing KV-cache memory requirements to one-quarter and SSD storage needs to one-eighth of the previous generation. The company also offers off-peak API rates at half the peak price.

In October, DeepSeek opened its Search and Fetch APIs to developers, pricing each at 5 yuan per 1,000 successful requests. Calls through DeepSeek API integrations with Claude Code and Codex are temporarily free, with per-request billing scheduled to begin on November 1.

These developments show DeepSeek addressing efficiency at both the model and infrastructure levels while expanding the tools available for AI agents. Whether the improvements translate into reliable performance and lower overall costs will depend on real-world workloads, deployment and maintenance.