DeepSeek · AI Programming

DeepSeek Open-Sources Agent Harness While Hiking API Prices Up to 12x

By 前端梦工厂 · Aug 14, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Any team running production workloads on DeepSeek's API faces an immediate cost recalculation—peak output is now 4.5× the old rate, and the steepest cache-hit tier rose 12×. The simultaneous Harness release signals that DeepSeek is betting on Agent-driven token volume to absorb the pricing shift, while giving developers a free, local-first Agent runtime to experiment with before the meter starts running.

Summary

Harness v0.1 ships as a local Node.js server with a Web UI, built on the principle that models, tools, sessions, sandboxes, storage, scheduling, and even the UI are all plugins. Four built-in modes range from a full tool suite to a minimal shell-and-file-editor setup for benchmarking. Custom model backends plug in directly.

The API pricing overhaul introduces a two-tier clock: peak hours (Beijing 9:00–12:00 and 14:00–18:00) cost double the off-peak rate. V4 Pro peak output hits 27 CNY per million tokens, 4.5× the old price. Cache-hit input for V4 Pro jumped from 0.025 CNY to 0.30 CNY—a 12× increase. Cache-hit savings are stark: on V4 Flash at peak, a hit costs roughly 96% less than a miss.

Releasing both on the same night looks like a coordinated play. Harness grows the Agent ecosystem and token consumption; the new pricing shifts compute costs onto heavy commercial usage while nudging batch workloads into cheaper off-peak windows.

Takeaways

— Harness is an MIT-licensed Agent runtime where models, tools, sessions, sandboxes, storage, scheduling, and the UI are all replaceable plugins.

— Starting Harness locally requires only Node.js and the command `npx @deepseek-ai/dsh web`, which serves a Web UI at localhost:3080.

— Four built-in modes ship with Harness: Standard (full tools), PTC (model-generated multi-tool calls), Minimal (shell + file editor for benchmarks), and Creative (runtime plugin experimentation).

— API pricing now follows a peak/off-peak schedule: peak is Beijing 9:00–12:00 and 14:00–18:00; off-peak is half the peak rate.

— V4 Pro peak output costs 27 CNY per million tokens, 4.5× the pre-adjustment price of 6 CNY.

— V4 Pro cache-hit input rose from 0.025 CNY to 0.30 CNY per million tokens, a 12× increase.

— Cache-hit savings are dramatic: on V4 Flash at peak, a hit costs about 96% less than a cache miss.

— New prices take effect August 17 at 00:00 Beijing time.

Conclusions

Releasing a free Agent runtime and a steep API price hike on the same night is a deliberate two-sided move: grow token-consuming Agent workloads with Harness, then monetize the resulting volume through the new pricing.

Peak-valley pricing functions as a demand-shaping lever, steering batch and non-urgent inference toward off-peak hours to smooth GPU utilization without adding capacity.

The 12× jump on cache-hit input for V4 Pro is the sharpest single-line increase and disproportionately penalizes workloads that cannot structure prompts for high cache-reuse.

Harness's 'everything is a plugin' architecture mirrors the modularity trend in Agent frameworks, but shipping it under MIT with a one-command local launcher lowers the barrier well below most alternatives.

Concepts & terms

Peak-Valley Pricing

A time-based pricing model where API costs are higher during defined peak hours and lower during off-peak hours, used here to incentivize shifting non-urgent workloads to cheaper time windows.

Cache Hit / Cache Miss

In LLM APIs, a cache hit occurs when a prompt prefix matches a previously computed key-value cache, allowing the model to skip recomputation and reducing cost. A cache miss requires full recomputation at the standard input rate.

Agent Runtime

The infrastructure layer that orchestrates an AI agent's loop—managing tool calls, session state, memory, sandboxes, and task scheduling—separate from the language model itself.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗