DeepSeek V4.1-Flash-1017 Outscores Larger Rivals on Agentic Benchmarks

DeepSeek has released V4.1-Flash-1017, an open-weight model that posts better benchmarks, finishes tasks faster, and produces shorter outputs than its predecessor — while once again cutting per-task costs against larger rivals.

The release lands in a market where agent frameworks, not chat windows, increasingly decide which models get deployed. Agentic workloads — coding assistants, browsing agents, multi-step tool use — punish models that ramble, stall, or burn tokens on dead-end reasoning. A model that completes the same task with fewer, shorter outputs lowers the bill at scale.

What does the new model actually deliver?

According to DeepSeek's announcement, V4.1-Flash-1017 improves on three axes simultaneously:

• Higher benchmark scores on agentic tasks, where the company says it beats larger rivals
• Faster task completion than its predecessor
• Shorter outputs, which directly reduces token consumption per task

The company frames the release as another step in its pattern of driving down per-task costs. That pattern is the reason DeepSeek commands attention beyond its benchmark results: each release has paired competitive performance with aggressive pricing.

The model ships as open-weight, meaning developers can download and run it on their own infrastructure rather than paying per token through a hosted API. That matters for enterprises with data-residency constraints and for labs that want to fine-tune or distill from the weights.

What are the new Search and Fetch APIs?

Alongside the model, DeepSeek is opening up its internal search engine through two new APIs:

• Search API — queries DeepSeek's search infrastructure directly
• Fetch API — retrieves web content for agent consumption

Both are priced at $0.75 per 1,000 calls. API users working in agent tools Claude Code and Codex get free access for now.

The pricing puts retrieval in reach for high-volume agent pipelines, where a browsing step can run on every user query. At $0.75 per 1,000 calls, a million-call retrieval workload costs $750 — low enough that search stops being the cost ceiling in most agent stacks.

The free access inside Claude Code and Codex signals where DeepSeek expects adoption: developer agents that need live web context. It also positions the company's search infrastructure as a product in its own right, not just an internal component of its models.

Why does this matter for the market?

The release continues the pricing pressure DeepSeek has applied since its earlier models disrupted assumptions about the cost of frontier-adjacent performance. Larger rivals — the hosted frontier models from Western labs — face a competitor that gives away the weights and charges fractions of a cent per retrieval call.

For buyers, the calculus is straightforward. An open-weight model that beats larger systems on agentic benchmarks, finishes faster, and emits fewer tokens per task shifts the economics of deploying agents from experiment to routine.

The shorter outputs carry a second-order benefit. Agents that produce concise intermediate steps complete multi-step workflows in less wall-clock time, which compounds across chained tasks. Speed and brevity are not cosmetic here; they are the cost lever.

What comes next?

Free access to the Search and Fetch APIs inside Claude Code and Codex is temporary, which suggests DeepSeek is collecting usage data before it settles on final commercial terms. Watch for the benchmark detail tables and independent evaluations — the announcement's claims of beating larger rivals will face the usual scrutiny once third parties run the weights.

For now, the combination of an open-weight model, sub-dollar retrieval pricing, and native placement in two widely used coding agents gives developers a concrete reason to re-price their agent stacks this week.