DeepSeek V4 Pro Goes GA: Open-Source Harness, Higher API Prices

DeepSeek just made a rare move in the middle of an AI price war: it open-sourced its agent software, then raised API prices in the same breath. On August 13, the lab officially moved DeepSeek V4 Pro (build 0813) out of testing, released Harness v0.1 — an MIT-licensed agent harness — and announced new peak/off-peak pricing that takes effect August 16.

Three announcements, one release. For developers on DeepSeek, all three hit directly.

V4 Pro: A Boost Aimed Squarely at Agents

The model name, parameter count, and 1M-token context window are unchanged — existing integrations keep running without tweaks. What changed is the brain inside, tuned for agentic workloads: debugging, self-correction, and multi-step tool calls.

By DeepSeek's own comparison table, Terminal Bench 2.1 jumped from 72.1 to 87.9, and DeepSWE from 12.8 to 62.7. On several agent benchmarks, V4 Pro 0813 beats Claude Opus 4.8. Take those numbers with the usual caveat — they're self-reported, and every vendor picks favorable benchmarks.

Independent numbers from Artificial Analysis are more measured: V4 Pro's Intelligence Index rose from 45 to 53, tying GLM-5.2 but still behind Muse Spark (57), Qwen 3.8 Max (58), Kimi K3 (60), and Claude Opus 5 (63). The upgrade is real, but it doesn't put V4 Pro in the lead pack. Notably, the 0813 weights aren't published yet — the April preview still sits on Hugging Face.

One detail that matters for agents: V4 Pro now natively supports the OpenAI Responses API with Codex integration, plus three reasoning-effort levels — low, high, max. DeepSeek recommends "high" for everyday agent work.

Harness: "Everything Is a Plugin"

Harness v0.1 (the dsh command) is the more interesting half of the announcement. It's an open-source agent harness pitched as an alternative to Claude Code or Codex — but architected to be open all the way down.

The core principle is "everything is a plugin": the model, tools, skills, sessions, sandbox, filesystem, loops, orchestration, and even the UI are swappable. It's built on Cordis, a newly released composable-plugin framework. Where Claude Code locks those components into one product, Harness lets you swap each layer.

The second differentiator is a continuous session log: every prompt, tool call, and result is recorded, and a run can be resumed, branched, or replayed. For debugging agents — notoriously hard to reproduce — that's genuinely useful.

Its current capabilities are enough to run a real coding agent: read repositories, edit files, run shell commands, search, plan, delegate to subagents, and enforce approval policies. DeepSeek is upfront that this is a Developer Preview — "THERE WILL BE COMPATIBILITY-BREAKING CHANGES" — so it's not production-ready yet. The project is led by Cui Tianyi, who joined DeepSeek from quant firm Jane Street in March 2026. An early-August call for beta testers drew 712 projects in three days.

Prices Up: Cache Hits Are the Pain Point

While the rest of the industry races to cut prices, DeepSeek is going the other way — and the sharpest increase lands on cache hits.

Starting August 16 at 16:00 UTC, DeepSeek drops flat pricing for peak/off-peak rates. Off-peak (outside the 1–4am and 6–10am UTC windows), input rises from $0.435 to $0.66 per million tokens, and output from $0.87 to $1.98. During peak hours, those double to $1.32 input and $3.96 output.

The heaviest hit is cache hits: from $0.003625 to $0.022 off-peak and $0.044 peak — a 6x to 12x jump. The cache discount shrinks from 1/120 to 1/30 of regular input price. For agents that repeatedly read the same files, that's the line item that balloons fastest.

The move partially undoes the price cut DeepSeek rolled out in May — and, per The Decoder, it lands as the company raises fresh capital.

What This Means for Developers

If you run agents on DeepSeek's API: recalculate your costs before August 16 — especially workloads that lean on cache hits. Prompt caching has been the standard way to keep read-heavy agents affordable; a 6–12x increase can upend that math.

If you want control over the agent layer: Harness is worth trying, especially if you need to swap models (DeepSeek, Anthropic, OpenAI, or custom endpoints) or want full session logs for debugging. The trade-off: it's young, with no GitHub-native PR workflow or hosted background agents like Claude Code yet.

If you only care about the model: V4 Pro 0813 is genuinely stronger, but independent numbers still place it mid-pack. Don't expect it to leapfrog Claude Opus 5.

The bigger signal is that DeepSeek no longer competes on model intelligence and token price alone. With Harness, it's moving into the layer that decides how models use tools, manipulate files, and run long workflows — the same territory Claude Code has come to dominate.

Key Takeaways

• V4 Pro build 0813 launched August 13; name, parameters, and 1M-token context unchanged.
• Terminal Bench 2.1 up 72.1 → 87.9; DeepSWE up 12.8 → 62.7 (DeepSeek self-reported).
• Artificial Analysis: Intelligence Index 45 → 53, tying GLM-5.2, behind Kimi K3 and Claude Opus 5.
• Harness v0.1 is open source (MIT), "everything is a plugin," launched via npx @deepseek-ai/dsh web.
• New pricing from August 16 16:00 UTC: peak/off-peak; cache hits up 6–12x.
• Install: npx @deepseek-ai/dsh web (Developer Preview).

---

In one day, DeepSeek handed the community an open-source agent tool and made its model more expensive to use. The two moves look contradictory but are really one strategy: shift the game from "cheapest" to "owning the entire agent loop." What's left to see is how much developers will pay for that software layer — and whether a young Harness can keep them when prices climb.

---

Content assisted by AI (Amy 🌸). Reviewed by the author.