DeepSeek Ships V4.1-Flash-1017: Open-Weight Model Beats Claude Opus 5 on Agentic Benchmarks While Slashing Per-Task Costs
DeepSeek released DeepSeek-V4.1-Flash-1017, an updated checkpoint of its V4.1-Flash model, as open weights under the MIT license on Hugging Face and ModelScope, alongside a Pareto improvement in performance and cost. The model scores 77.0 on DeepSWE v1.1, up from 74.2 for the previous version, surpassing Claude Opus 5's 74.0 on the same benchmark. At the same time, the company opened its Search and Fetch APIs to all developers at $0.75 per 1,000 successful requests, keeping the new model's effective cost under $0.20 per agentic task.
Benchmark Gains Through Reinforcement Learning With Self-Feedback
The updated checkpoint was trained with reinforcement learning using self-feedback, which DeepSeek credits for faster task completion, shorter outputs, and stronger agentic and coding abilities. The model activates 8B parameters per token during prefill and 16B during decode, supporting a 1M-token context window and up to 384K output tokens. DeepSeek positions the release as a Pareto improvement: higher scores with lower latency and cost, rather than a trade-off between the two.
Search and Fetch APIs Opened to All Developers
Alongside the model, DeepSeek made its in-house search engine — which it says handles hundreds of millions of daily requests — available through public Search and Fetch APIs priced at $0.75 per 1,000 successful calls per endpoint. Access is free through server-side tools in DeepSeek Harness, Claude Code, and Codex until November 1, 2026, after which per-request billing begins. This gives agent builders retrieval infrastructure beyond DeepSeek's earlier preview, directly competing with paid search toolchains from US providers.
Pressure on Closed-Weight Pricing
By delivering competitive agentic and coding performance at under $0.20 per task, DeepSeek is increasing pressure on OpenAI and Anthropic to justify premium API pricing. The MIT-licensed weights, FP8 support, and wide distribution through OpenRouter and major registries make the model a first-class option for production use, especially for teams with data-sovereignty requirements that prefer self-hosting. Analysts note the release intensifies the open-weight speed race against Meta's Llama family, Qwen, Mistral, and Google's open releases.
Sources: The Decoder, Hugging Face, api-docs.deepseek.com, Hacker News, MarkTechPost, threerouter
Top comments (0)