DeepSeek Releases V4.1-Flash-1017: Lower Inference Costs and Public Search API
DeepSeek updated its V4.1 Flash model to V4.1-Flash-1017 on October 17, achieving a 77.0% DeepSWE v1.1 pass rate, opening its search API, and releasing weights
On October 17, 2026, DeepSeek officially announced the V4.1-Flash-1017 update across web, application, and API channels. While official technical documentation categorizes this release as a minor update following the September 10 launch of V4.1 Flash, the accompanying data highlights tangible gains in coding benchmark efficiency and lower per-task execution costs.
Benchmark Performance and Cost Optimization
In official tests across 11 benchmarks, V4.1-Flash-1017 posted a 77.0% pass rate at top tier on DeepSWE v1.1, surpassing Claude Opus 5 (74.0%), GPT-5.6-Sol (73.0%), Kimi K3 (67.5%), and GLM 5.3 (66.9%). DeepSeek noted reduced runtimes and output lengths lowered top-tier DeepSWE v1.1 task costs from 3.3 yuan (~$0.49) to 1.3 yuan ($0.19). It also led Automation-Bench v1.0.6 (60.5) and Apex Bench v1.1 (72.1). However, Claude Opus 5 maintained leads on HLE (56.3), Terminal-Bench 4.0 (51.8), ProgramBench (37.0), and NL2Repo-Bench (75.3).
Search API Access and Open-Source Weights
DeepSeek opened its proprietary search engine to the public, introducing Search and Fetch API endpoints at 5 yuan (~$0.75) per 1,000 calls. The engine indexes hundreds of billions of web pages and filters SEO spam and low-quality AI content according to vendor claims. Furthermore, model weights with 763B parameters and a 1M-token context window were released on HuggingFace and ModelScope under the MIT license. API endpoints support OpenAI (https://api.deepseek.com) and Anthropic (https://api.deepseek.com/anthropic) formats at existing deepseek-flash pricing.
Practical Impact, Caveats, and Limitations
For teams deploying coding agents, the update lowers task costs through reduced token usage without base price increases. However, essential caveats remain. All reported metrics, pass rates, and cost calculations are based on DeepSeek's initial release documentation rather than third-party audits. Search filtering capabilities also reflect vendor claims. With Claude Opus 5 retaining its lead on terminal tasks and complex reasoning, teams should conduct workload-specific validation prior to production migration.