DeepSeek V4.1-Flash-1017 Tops DeepSWE, Opens $0.75 Search API

DeepSeek released V4.1-Flash-1017 on October 17, a 552B-parameter sparse MoE model with 8B active parameters during prefill and 16B during decoding, which scores 77.0 on the DeepSWE v1.1 coding benchmark ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) at under $0.20 per task. DeepSeek also opened standalone Search and Fetch APIs at $0.75 per 1,000 calls, roughly one-tenth of Tavily's $8.00 and 18x cheaper than Google Gemini grounding at $14.00, with billing starting October 31 at 16:00 UTC after a free period. The APIs carry a 5 QPS cap, a 128-character query limit, and process data on servers in the People's Republic of China, which DeepSeek's privacy policy states is where data is collected, processed and stored.

On October 17, DeepSeek shipped two things at once: V4.1-Flash-1017, a model update that tops Claude Opus 5 on the DeepSWE coding benchmark, and Search and Fetch APIs priced at $0.75 per 1,000 calls — roughly one-tenth of what Tavily charges. Either story would stand on its own. Together they signal that DeepSeek isn’t just building cheaper models anymore. It’s building the retrieval infrastructure for agents. What Changed in V4.1-Flash-1017 The model card bills V4.1-Flash-1017 as a “minor update” to the existing Flash tier. The benchmark numbers tell a different story. On DeepSWE v1.1 https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/ — a rigorous software engineering benchmark that measures multi-step code repair and agent-driven task completion — the new build scores 77.0, ahead of Claude Opus 5 74.0 and GPT-5.6 Sol 73.0 . Per-task cost on that same benchmark: under $0.20. That’s the headline number. The others are worth knowing too. Terminal-Bench 4.0 jumps from 31.2 to 42.3, ProgramBench climbs from 20.3 to 29.1, and Agents’ Last Exam reaches 35.4 versus GPT-5.6 Sol’s 26.7. One honest caveat: Claude Opus 5 holds a ~10-point lead on Terminal-Bench 4.0 at 51.8, so this isn’t a clean sweep of every agentic task. DeepSWE is where the Flash tier genuinely pulls ahead. The architecture is a 552B-parameter sparse MoE with 8B active parameters per token during prefill and 16B during decoding. Context window is 1M tokens. Weights ship under MIT on Hugging Face https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 , compatible with vLLM, SGLang, TensorRT-LLM, and llama.cpp. The API model name stays deepseek-flash — no migration required if you’re already using the V4.1 Flash endpoint. The Search and Fetch APIs The more interesting release for agent builders may not be the model at all. DeepSeek opened its internal search infrastructure as two standalone APIs: Search returns ranked results for a query; Fetch converts any URL into clean Markdown. Both authenticate with your existing DeepSeek API key. curl https://api.deepseek.com/v1/tools/search -H "Authorization: Bearer <your-key " -H "Content-Type: application/json" -d '{"query": "your query here", "count": 5, "need content": true}' On price, they’re not even close to the competition: | Provider | Price per 1,000 calls | |---|---| | DeepSeek Search / Fetch | $0.75 | | Perplexity API | $1.00 | | Exa | $4.00 | | Brave Search API | $5.00 | | Tavily basic | $8.00 | | Google Gemini grounding | $14.00 | DeepSeek is 7x cheaper than Brave, 10x cheaper than Tavily, and 18x cheaper than Google’s Gemini grounding. For an agent that runs 50,000 searches per month, the difference between DeepSeek and Tavily is over $360 monthly. That’s real money at scale. Full API documentation is available at DeepSeek’s Search and Fetch API docs https://api-docs.deepseek.com/guides/search/ . The Catches The pricing is real, but so are the constraints. Read these before you integrate. 5 QPS cap. Five requests per second per account. If your agent fans out parallel search calls, you’ll hit this fast. There’s no enterprise tier announced yet. Data routes to China. Queries and fetched URLs are processed on DeepSeek’s servers in the PRC. Their privacy policy explicitly states data is “collected, processed and stored in the People’s Republic of China.” If your agent handles anything regulated — HIPAA records, ITAR-controlled data, anything under EU GDPR with explicit data residency requirements — this is a blocker, not a footnote. 128-character query limit. Search queries cap at 128 Unicode characters. Long-form or highly specific queries need to be condensed before hitting the API. Billing starts October 31. Both APIs are currently free for users of Claude Code, Codex, and DeepSeek Harness. That free period ends October 31 at 16:00 UTC. Build the dependency now, pay later — it’s a textbook land-grab window. The Bigger Picture DeepSeek entering the retrieval layer is the part of this announcement that the model benchmarks tend to overshadow. Tavily, Exa, and Brave have built real businesses charging developers for web search access. DeepSeek just dropped into that market https://thenewstack.io/deepseek-web-search-api/ at prices none of them can easily match without restructuring their cost base. The play here is obvious: build agentic pipelines that rely on DeepSeek for both inference and retrieval, then the switching cost becomes the full stack. For developers building internal tools or handling non-sensitive data, that combination — fast model, cheap search, single API key — is genuinely compelling. For anyone handling regulated data or operating in jurisdictions with data residency rules, self-hosting the model via the open weights sidesteps the China routing issue for inference, though not for the Search and Fetch APIs themselves. At $0.75 per 1,000 calls, the Search API is worth testing in any stack where Tavily or Brave is currently the cost you’re most likely to cut. Just know what you’re sending before you send it.