Skip to main content

DeepSeek-V3.1 Released: The First Step Toward the Agent Era (Quick Guide + Key Insights)

AIDeepSeekV3_1AI_AgentFunctionCallingLargeLanguageModelAIProductivityHybridReasoning128KContextAbout 2 min

DeepSeek-V3.1 Released: The First Step Toward the Agent Era (Quick Guide + Key Insights)

On August 21, DeepSeek officially released DeepSeek-V3.1. This generation introduces major upgrades in hybrid reasoning (thinking/non-thinking modes), efficiency, and Agent toolchain capabilities. The update is already live in the official App and web client (with a new "Deep Thinking" toggle). For API usage, deepseek-chat corresponds to non-thinking mode, while deepseek-reasoner is the thinking mode, both now supporting a 128K context window. In addition, strict Function Calling (Beta) and Anthropic API compatibility are introduced, making it easier to integrate with ecosystems such as Claude Code.

Key Changes in V3.1

  • Hybrid reasoning architecture: One model, two modesโ€”switch freely between โ€œthinkingโ€ and โ€œnon-thinking.โ€
  • Efficiency gains: 20%โ€“50% fewer tokens on average in thinking mode while maintaining performance comparable to R1-0528.
  • Stronger Agent capabilities: Significant improvements in coding, terminal tasks, and multi-step search.
  • Developer-friendly features: 128K context window, strict Function Calling (Beta), Anthropic API compatibility, smoother and more scalable API access.

Public benchmarks show strong gains in both programming agents and search agents:

  • SWE Verified (Agent mode): V3.1 66.0 vs R1-0528 44.6
  • Terminal-bench (Terminus 1): V3.1 31.3 vs R1-0528 5.7
  • BrowseComp (multi-step search): V3.1-Think 30.0 vs R1-0528 8.9
  • HLE (expert-level multi-domain challenges): V3.1-Think 29.8 vs R1-0528 24.8

This signals more reliable tool usage and deeper reasoning for search-driven tasks.

Reasoning Efficiency: Shorter Chains, Same Accuracy

Through chain-of-thought compression training, V3.1-Think reduces 20%โ€“50% tokens per output while keeping results on par with R1-0528. For example:

  • AIME 2025: 88.4 vs 87.5
  • GPQA: 80.1 vs 81
  • liveCodeBench: 74.8 vs 73.3

In practice, this means lower latency and cost while preserving reasoning accuracy.

5 Things Developers Must Know

  1. Choosing the right model
    Use deepseek-chat for fast, lightweight tasks; use deepseek-reasoner for complex reasoning. Both support 128K context.

  2. Strict Function Calling (Beta)
    Set base_url=https://api.deepseek.com/beta, add strict: true in the schema. The server will enforce schema validation, ensuring function outputs strictly conform.

  3. Anthropic API compatibility
    Configure ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic to call DeepSeek models directly in Claude Code and other Anthropic-based SDKs.

  4. Open-source and precision updates

    • Released models: Both Base and post-trained weights (available on Hugging Face & ModelScope).
    • Extended pretraining: +840B tokens on top of V3 for longer context and stability.
    • Precision: UE8M0 FP8 Scale, with updated tokenizer and chat template (migration required).

  5. Pricing & policy changes
    From September 6, 2025 (Beijing Time), a new pricing plan will take effect, and nighttime discounts will be removed. Until then, current pricing applies.

Quick Start in 15 Minutes

  • Pick your mode: Use deepseek-chat for efficiency, deepseek-reasoner for deep reasoning with 128K.
  • Enable strict FC (optional): Define tools with strict schema to avoid invalid calls.
  • Seamless integration: Switch base URL in Claude Code/Anthropic SDK to instantly adopt.
  • Retrieval agents: Search โ†’ filter โ†’ synthesize workflows.
  • DevOps & coding assistants: Code fixes, patches, scripting, command-line tasks.
  • Long conversations with cost control: 128K + compressed reasoning = context depth with reduced expense.

FAQ

Q1: How can I switch between thinking and non-thinking modes?
In the App/web, click the "Deep Thinking" button; in the API, use different model names or templates.

Q2: Whatโ€™s the value of strict Function Calling?
It ensures schema-compliant outputs, reducing formatting errorsโ€”ideal for production-ready Agents.

Q3: Can I integrate with Anthropic/Claude Code?
Yes. Just set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and API Key.

Q4: Whatโ€™s new in V3.1โ€™s open-source release?
Both Base and post-trained models are open-sourced, with +840B tokens of extended training, UE8M0 FP8 precision, and an updated tokenizer/template.

Q5: When will the new pricing take effect?
September 6, 2025 (Beijing Time). Nighttime discounts will be removed, but current rates remain until then.

Find Cheap and Best servers.