DeepSeek V4 Pro Launches with Major Agent Boost and Hints at Upcoming Harness

1M Context Retained, Agent Capabilities Keep Rising

The official V4 Pro release continues the V4 series with a 1 M token context window and a maximum output length of 384 K tokens, supporting both thinking and non‑thinking modes. JSON Output, Tool Calls, Responses API, and Anthropic API are all available, making the model especially useful for AI programming scenarios where large codebases, requirement documents, and conversation histories need to be processed.

DeepSeek’s published agent evaluation shows clear improvements over the April preview:

Terminal Bench 2.1: 72.1 → 87.9

NL2Repo: 38.5 → 61.5

DeepSWE: 12.8 → 62.7

AutomationBench (Public): 12.8 → 31.8

DSBench‑Hard: 32.1 → 67.2

These benchmark gains indicate that DeepSeek is strengthening long‑task handling, tool invocation, and software‑development execution capabilities, though real‑world projects may not see identical improvements.

It’s Not the Best at Every Metric

When public scores are combined with those of GPT‑5.6 Sol, Claude Opus 5, and other models, DeepSeek V4 Pro does not dominate every column. Closed‑source leaders still hold advantages in HLE, DeepSWE, and Toolathlon‑Verified, while the model approaches top‑tier performance on Terminal Bench 2.1.

The author argues that model selection should consider cost, context length, API compatibility, and task stability rather than chasing a single champion. V4 Pro 0813’s appeal lies in bundling usable agent abilities with relatively low API costs.

Pricing Remains Stable—for Now

Current API pricing (CNY) is:

1 M input tokens, cache hit: 0.025 ¥

1 M input tokens, cache miss: 3 ¥

1 M output tokens: 6 ¥

The endpoint URL stays the same (https://api.deepseek.com), so existing applications using deepseek-v4-pro can switch to the new version without code changes. A Node.js example demonstrates a simple call:

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.DEEPSEEK_API_KEY,
  baseURL: "https://api.deepseek.com",
});

const response = await client.responses.create({
  model: "deepseek-v4-pro",
  input: "分析这个项目的测试失败原因，并给出修复方案。",
});

console.log(response.output_text);

DeepSeek has indicated on its pricing page that a substantial API price increase is planned soon, so teams with stable usage should re‑evaluate budgets rather than assuming the current rates will persist.

DeepSeek Harness May Arrive Soon

Community screenshots suggest that DeepSeek Harness (DSH) will end its closed beta on August 13 and enter public testing, with possible migration of internal plugins to developers’ own accounts.

Although no official announcement exists yet, the V4 Flash update log mentioned that the public Code Agent benchmark used a “DeepSeek Harness minimal mode (coming soon)”. If Harness is released, V4 Pro 0813 would pair a powerful inference model with a workflow layer handling task decomposition, tool scheduling, context management, and result aggregation, delivering a ready‑to‑use agent system for developers.