# DeepSeek Just Attacked Claude Code: Harness + V4 Pro Video: https://www.youtube.com/watch?v=WPlOUcQFlKU Article: https://aidive.dev/videos/deepseek-harness-v4-pro/ Published: 2026-08-14 ## Chapters - [0:00](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=0s) DeepSeek strikes twice in one day - [1:03](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=63s) What an agent harness actually is - [2:03](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=123s) Everything is a plugin: inside dsh - [4:11](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=251s) V4 Pro: the engine behind the harness - [6:04](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=364s) The pricing gap with Claude - [7:42](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=462s) Can it replace Claude Code? - [8:59](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=539s) The traps before you migrate - [9:50](https://www.youtube.com/watch?v=WPlOUcQFlKU&t=590s) So, migrate or not? ## TL;DR - On August 13th, 2026, DeepSeek shipped DeepSeek Harness (dsh), an MIT-licensed open-source coding agent aimed directly at Claude Code, and moved DeepSeek V4 Pro to general availability the same day. The combo of a harness plus a model is the same formula that made Claude Code dominant, which is why Bloomberg framed it as a direct challenge to Anthropic. - dsh's core idea is 'everything is a plugin': models, tools, sandboxes, storage, loops, and even the UI are swappable in config, on top of an append-only session log you can replay event by event. - V4 Pro claims ~80.6% on SWE-bench Verified (vs 80.8% for Claude Opus 4.6) with a 1M-token context window — but every benchmark is self-reported, with no third-party reproduction yet. - At launch pricing, V4 Pro is ~11x cheaper than Claude Opus 5 on input and ~28x on output; from August 16th, peak/off-peak billing raises output up to $3.96/M at peak — still ~4x cheaper than Opus 5. - Verdict: tooling builders should try dsh now, cost-cutters should test V4 Pro on side projects at the post-August-16 rates, and daily Claude Code users should stay put — dsh is a 0.1 developer preview that promises breaking changes. ## DeepSeek strikes twice in one day On August 13th, 2026, DeepSeek shipped two products at once. DeepSeek Harness is an open-source coding agent under an MIT license that goes straight after Claude Code — Bloomberg described it in plain terms as a challenge to Anthropic. The same day, DeepSeek V4 Pro (build 0813) hit general availability, with a one-million-token context window priced under half a dollar per million input tokens. The market noticed immediately. The Hacker News post for the model cleared 990 points in a single day, and the Harness repository already sits at over 21,000 GitHub stars. A tool plus a model shipped on the same day is exactly the formula that made Claude Code what it is. We installed DeepSeek Harness with one npx command, read the docs, and went through every number DeepSeek published. This article covers the everything-is-a-plugin architecture, what V4 Pro is worth against Claude on benchmarks and pricing, and who should migrate versus wait. ## What an agent harness actually is A harness is all the machinery wrapped around a model so it can actually work on your computer: the loop that chains the calls, the tools the model is allowed to use, the context management, the permissions, and the interface. The model decides, the harness executes — that short sentence is the whole division of labor. When you type an instruction into Claude Code, the model never touches your disk itself. It asks the harness to read a file, run a command, or launch a test; the harness checks whether that's allowed, executes it, and hands the result back. Every coding agent works this way. Claude Code is Anthropic's harness wired to the Claude models, and Codex is OpenAI's harness wired to GPT. DeepSeek never had an official one — its models ran inside other people's tools. That's what August 13th changes: DeepSeek is done selling just the engine and is now selling the whole car. ## Everything is a plugin: inside dsh DeepSeek Harness — shortened to dsh in the docs — is built on one central idea: everything is a plugin. That's not a marketing slogan, it's the actual architecture. Models are plugins. Tools are plugins. Skills, sessions, sandboxes, storage, loops, scheduling, and even the interface: all plugins. The official docs state you can select, swap, or extend any capability in configuration without changing the harness source code. Under the hood it rests on Cordis, a kernel DeepSeek describes in a research paper on composability. The project has real weight behind it: over 12,000 commits, a Discord server, and a dedicated `dsh-plugin` tag on GitHub where community plugins are already showing up. The practical consequence is that the harness has no frozen core — if you want a different sandbox, different session storage, or a different model, you change the config, not the code. Installation is one command: `npx @deepseek-ai/dsh web` opens a web interface locally on port 3080. No account required to start, no installer — the npm package is enough. If you'd rather build from source, the whole repo clones and compiles with pnpm in four commands. dsh ships four runtime modes, and this is where the architecture bet becomes visible: - **Standard mode** is the full coding agent: file editing, shell, search, and planning — the direct equivalent of Claude Code. - **Code mode** has the model generate TypeScript that orchestrates multi-step operations itself instead of chaining tool calls one by one, which cuts round trips, latency, and tokens on long tasks. - **Minimal mode** strips the agent down to bash and a file editor, mostly for benchmarking the raw model. - **Creator mode** is for building your own presets, with runtime inspection and live plugin experiments. The last piece is traceability. Everything the model sees goes into an append-only session log — you can only add to the end, never rewrite history. That log is the harness's source of truth, and you can resume, fork, search, or replay any session from it. When your agent goes off the rails on tool call forty-two, you don't dig through a transcript; you replay the session to the tipping point and see exactly what the model saw. That's operating-system thinking, not another API wrapper. ## V4 Pro: the engine behind the harness DeepSeek V4 Pro, build 0813, is the model leaving preview and hitting general availability, already runnable through the DeepSeek API or OpenRouter. It's a mixture-of-experts model, meaning only a fraction of the network activates per token, so you get big-model capability at small-model compute cost. The spec sheet: a one-million-token context window, up to 384,000 output tokens, tool calls, and JSON output. One detail counts double in this story: the DeepSeek API advertises Anthropic API compatibility. A tool written to talk to Claude can talk to V4 Pro by swapping a base URL and a key. On benchmarks, DeepSeek claims scores near the top: | Benchmark | DeepSeek V4 Pro | Claude | |---|---|---| | SWE-bench Verified | ~80.6% (Max variant) | 80.8% (Opus 4.6) | | Terminal Bench 2.1 | claimed win | Opus 4.8 | | Artificial Analysis index | 53 | 63 (Opus 5) | | Speed (tokens/sec) | 83 | 52 | But one detail changes how you read every one of those numbers: not a single score has been reproduced by a third party yet. Everything in circulation traces back to DeepSeek or to articles copying DeepSeek. Simon Willison points out the benchmarks first circulated in DeepSeek's official WeChat group, got copied into a Reddit post that moderators deleted, and ended up as an ASCII table on Hacker News. He also notes there's no official announcement page, and open weights aren't confirmed — though April's V4 Pro and July's V4 Flash both ended up on Hugging Face, so a release looks likely. His pelican test at least showed something unusual: three reasoning levels produced three radically different pelicans. The model is probably very good; right now, DeepSeek is the only one saying so. ## The pricing gap with Claude Price is where DeepSeek hits hardest. At launch, V4 Pro costs $0.435 per million input tokens and $0.87 per million output tokens through the DeepSeek API. | Model | Input ($/M tokens) | Output ($/M tokens) | |---|---|---| | DeepSeek V4 Pro (launch rate) | $0.435 | $0.87 | | Claude Sonnet 5 | $2 | $10 | | Claude Opus 5 | $5 | $25 | | Claude Fable 5 | $10 | $50 | At the launch rate, V4 Pro runs about 11x cheaper than Opus 5 on input and 28x cheaper on output. Even Sonnet 5, Anthropic's budget option, stays four to eleven times more expensive. For a typical coding agent session of 50,000 tokens in and 15,000 out, Opus 5 costs about $0.62; V4 Pro at the launch rate costs about $0.035. The hidden line item for agents is cache, because a harness resends the same context every loop turn: a cache hit costs $0.50 per million on Opus 5 and about $0.0036 per million on DeepSeek. Except this floor price has an expiration date, and it's close. Starting August 16th — three days after launch — the API moves to peak and off-peak billing: | Rate period | Input ($/M) | Output ($/M) | |---|---|---| | Launch (until Aug 16) | $0.435 | $0.87 | | Off-peak | $0.66 | $1.98 | | Peak (1–4 AM and 6–10 AM UTC) | $1.32 | $3.96 | The price everyone is comparing this week rises by at least half, and peaks at over quadruple on output. Even at the worst rate, V4 Pro stays almost four times cheaper than Opus 5, so the argument survives — but the number going around Hacker News is a promo ending in three days. The schedule also matters for agents: peak hours track the Chinese workday, so overnight crons can hit off-peak rates while afternoon coding eats peak windows. ## Can it replace Claude Code? On architecture, dsh has arguments Claude Code doesn't. When your agent derails, Claude Code lets you dig through the transcript; dsh lets you replay the session event by event. The plugin system reaches further than Anthropic's skills and MCP servers, because it also covers the sandbox, the storage, and the loop itself — in Claude Code you extend the agent, in dsh you can recompose it. The harness also takes models other than DeepSeek's, where Claude Code stays locked to Claude unless you route through unofficial proxies. On the other side, Claude Code has maturity, and it weighs a lot: months of production across tens of thousands of teams, a huge ecosystem of skills, hooks, and integrations, models whose scores independent evaluators verify, and subscription access that shields you from billing surprises, where dsh bills every token. dsh is a developer preview at version 0.1, and the README says in capital letters that there will be compatibility-breaking changes. But the Anthropic API compatibility means you don't actually have to pick a side: keep your harness and point workflows at V4 Pro, or try dsh with the model you already use. dsh is the most credible challenger Claude Code has faced — and it's a 0.1 challenger against a sitting champion. ## The traps before you migrate There are three limits to weigh before uninstalling anything. First, dsh is a developer preview that breaks compatibility without notice, so a plugin you write this week may not load next month. Building your daily workflow on it today means agreeing to repair it regularly. Second, every V4 Pro benchmark is self-reported. No independent reproduction, no announcement page, numbers that surfaced in a WeChat group. They may well be accurate, but nobody outside DeepSeek knows that yet. Third, the price making the buzz is on borrowed time. From August 16th, time-slot billing pushes output up to $3.96 per million at peak — four and a half times launch-day pricing. dsh gives you total freedom over your agent's architecture and almost no guarantee about what will still be true in a month. If your agent stack runs your business, this is not the week to move it. ## So, migrate or not? Our read after a day inside it comes down to three profiles. If you build harnesses or agent tooling, install dsh this week: the all-plugin architecture and the replayable log are the two most interesting ideas around right now, and one npx command gets you an opinion. If you just want a smaller API bill on agent tasks, test V4 Pro on a side project — but run the math on the August 16th rates, and wait for independent confirmation of the benchmark scores before rerouting production. If Claude Code runs your daily work, keep it. Nothing in this launch justifies migrating a tool that works to a version 0.1 that promises breakage. The good news is that this attack forces everyone to move: whatever agent you're on, its price now has a real chance of dropping. ## FAQ ### What is DeepSeek Harness (dsh)? DeepSeek Harness, shortened to dsh, is an open-source coding agent released by DeepSeek on August 13th, 2026 under an MIT license. It is DeepSeek's equivalent of Claude Code, built on an everything-is-a-plugin architecture where models, tools, sandboxes, storage, and even the interface can be swapped in configuration. ### What is an agent harness? A harness is the machinery wrapped around an AI model so it can work on your computer: the loop that chains calls, the tools the model may use, context management, permissions, and the interface. The model decides, the harness executes — Claude Code is Anthropic's harness for Claude, and Codex is OpenAI's harness for GPT. ### How do you install DeepSeek Harness? Run `npx @deepseek-ai/dsh web` and a local web interface opens on port 3080. No account or installer is required to start, and the repo can also be built from source with pnpm in four commands. ### Is DeepSeek V4 Pro better than Claude? DeepSeek claims near-parity: about 80.6% on SWE-bench Verified versus 80.8% for Claude Opus 4.6, and 83 tokens/sec versus 52 for speed, though Opus 5 leads 63 to 53 on the Artificial Analysis index. None of these scores has been independently reproduced yet — all published numbers trace back to DeepSeek itself. ### How much cheaper is DeepSeek V4 Pro than Claude? At launch pricing ($0.435/M input, $0.87/M output), V4 Pro is about 11x cheaper than Claude Opus 5 on input and 28x on output. From August 16th, 2026, peak/off-peak billing raises prices to $0.66–$1.32 input and $1.98–$3.96 output, which still leaves it roughly 4x cheaper than Opus 5 at the worst rate. ### Should I switch from Claude Code to DeepSeek Harness? Not for daily production work yet. dsh is a 0.1 developer preview whose README warns of compatibility-breaking changes, and V4 Pro's benchmarks are unverified. It is worth installing if you build agent tooling, and V4 Pro is worth testing on side projects thanks to Anthropic API compatibility.