DeepSeek's Coding Agent vs Claude Code: Harness and Super-Agent Against Anthropic's Agent on Features, Models, Benchmarks, Price, Availability, and Safety, in One Table

Features, price, and availability in one table

The form-factor difference, which the headlines skip

Claude Code is a product you install and drive. Open a terminal, type claude, and Opus 5.5 reads the repository, plans, edits, runs tests, and commits while you watch, with a permission classifier reviewing risky actions in the background. The same agent is now in VS Code and JetBrains, a desktop app, claude.ai/code in the browser and the mobile app, Slack, a Chrome extension, and GitHub Actions. Everything is Anthropic's: the binary 'must not be modified,' the license is 'all rights reserved,' and the gateway docs say Anthropic 'doesn't support routing Claude Code to non-Claude models through any gateway.'

DeepSeek Harness is a runtime you host. One command starts a web UI; the desktop app covers Apple-silicon Macs and Windows. Inside it you get sessions that look a lot like Claude Code's (terminal, subagents, skills, background tasks, scheduled tasks), plus things Claude Code doesn't do in the same place (documents, spreadsheets, slides, 'Creator mode' that writes a new plugin in chat). The whole thing is plugins on a framework called Cordis, MIT-licensed, and DeepSeek's README says in capitals that 'there will be compatibility-breaking changes.' Super-agent sits on top as a bundle: a Router that triages a repository's events and Workers that take tickets, with a philosophy of taking 'as little human attention as possible.' Claude Code's nearest equivalents are cloud scheduled tasks triggered by GitHub events, GitHub Code Review, and Channels; none of them monitors a team chat for bugs, and none of them is described as a prototype.

Models and benchmarks: Opus 5.5 is a tier above, and DeepSeek knows it

Anthropic's Opus 5.5 announcement reports 66.4% on Terminal-Bench 4.0 and 89.9% on SWE-bench Pro, and Artificial Analysis's Coding Agent Index puts Claude Code with Opus 5.5 first at 62, ahead of Gemini 4 Carbon in Antigravity at 59. DeepSeek's own V4.1 Flash 1017 card reports 42.3% on Terminal-Bench 4.0, and it ran that number inside the Claude Code harness, which tells you which tool DeepSeek benchmarks against. On the one independent board with both stacks, Collinear's CWE-bench, Opus 5.5 in Claude Code scores 67% at $0.79 per rollout and V4.1 Flash scores 55% at $0.09. That is the trade in one row: twelve points of capability for about nine times the price per attempt. Flash is a 763-billion-parameter model with 8 to 16 billion active; it is not trying to be Opus. V4.1 Pro, which would be, has no date.

The honest version of 'DeepSeek vs Claude Code' on quality is therefore 'DeepSeek Harness running Opus 5.5 vs Claude Code running Opus 5.5,' and that comparison nobody has published. Harness accepts custom endpoints, and DeepSeek's card for the September V4.1 Flash claims its own minimal scaffold beat the Claude Code, Codex, and OpenCode scaffolds on the same model, which is a harness claim, not a model claim. We'll test it; until then the only measured gap is the model gap, and it is large.

Price: dollars versus cents, with two catches

A Claude Pro subscriber pays $20 and gets Claude Code with a five-hour session limit and a weekly limit shared with the Claude apps; heavy users go to Max at $100 or Ultra at $200 (Ultra is the renamed Max 20x, with no five-hour limit when verified). On the API, Opus 5.5 is $4 in and $20 out per million tokens, Sonnet 5.5 $2 and $10. DeepSeek Harness costs nothing; V4.1 Flash on the API is $0.15 in and $0.60 out per million off-peak, and a cache hit is $0.003, which is why Super-agent's test could read 23 million cached tokens a run without anyone flinching. Catch one: DeepSeek's prices double during Beijing business hours (01:00–04:00 and 06:00–10:00 UTC on weekdays). Catch two: cheap tokens on a smaller model can mean more tokens; Artificial Analysis calls V4.1 Flash 'very verbose' at 89,000 output tokens per index task. Run your own workload through our token counter with both models' rates before assuming the ratio.

Availability: the map decides for a lot of people

Claude Code is not available where DeepSeek's home users are. Anthropic's supported-countries list omits mainland China, Hong Kong, and Macau, and since September 2025 Anthropic bars sales to companies more than 50% owned from unsupported regions 'regardless of where they operate.' DeepSeek Harness is a 'public preview worldwide' and the code is MIT, so a developer in Shanghai or a company that can't pass Anthropic's ownership test has one of these two options, not two. The reverse constraint is softer but real: US federal civilian agencies were reportedly told in October to remove DeepSeek software, and many Western enterprises won't send code to a Chinese API, which is where self-hosting Harness on a different model becomes the pitch.

Safety and trust

Claude Code's model is permissions plus review: a classifier-backed auto mode, a plan mode that proposes before acting, checkpoints you can rewind, and hooks that let you block or audit any of 33 lifecycle events. Its source is closed, which cuts both ways; a packaging mistake in March leaked it, and last week a trojanized copy of its VS Code extension went through Cursor's marketplace, a reminder that a closed binary is still a supply chain. DeepSeek's model is openness plus warnings: you can read every line, and the safety file tells you it 'has not undergone a security audit and must not be treated as secure or production-ready,' recommends 'a disposable virtual machine, container, or dedicated environment,' and Super-agent's workers 'run commands without approval as the operator's OS user' behind wrappers that are 'not a security boundary.' Neither posture is wrong; they ask different things of you.

Which should you use?

• You write code for a living and can pay $20–200 a month: Claude Code. The model gap is the product gap, and the surfaces, permissions, and polish are a year ahead.
• You want an agent that owns a repository while you sleep: try Super-agent on a throwaway repository in observe mode, and compare it with Claude Code's cloud scheduled tasks and Copilot's coding agent. Nobody has measured what Super-agent merges yet, including DeepSeek.
• You can't use Anthropic, by geography or ownership: DeepSeek Harness is the most complete open agent runtime available, and it runs on V4.1 Flash for cents. Budget engineering time for breaking changes.
• You want to own the harness and bring any model: Harness, with a custom endpoint, or Aider if you want something smaller and stable. Claude Code will not run non-Anthropic models.
• You're choosing an open-weight model rather than a tool: our open-source ranking and the Kimi K3.1 Pro comparison cover the frontier open models; DeepSeek's weights are MIT but 284 billion parameters and up, so you'll rent them.

What changes in our rankings

Claude Code holds #2 in AI code editors at 4.8. DeepSeek Harness joins that category's 'also considered' list as a provisional entry we have not scored hands-on, marked as such, because a developer preview that promises breaking changes and declines a security audit is not something we rank against shipped products on documentation alone. Super-agent is not ranked; it isn't a product by its author's description. Both go on our test list, and the Claude Code review's pricing now reads Max and Ultra, since the Max 20x tier no longer exists.

Sources and what's unconfirmed

• Primary: Claude Code docs, pricing, and supported-countries pages; Anthropic's Opus 5.5 announcement and system card; the S-1; DeepSeek's Harness repository, safety file, official page, and release notes; the Super-agent repository and design documents; DeepSeek's API pricing page and V4.1 Flash model cards; Collinear's CWE-bench board; Artificial Analysis's Coding Agent Index.
• Secondary: Bloomberg's October 22 and November 1 reports (the $15 billion round and the Super-agent story); Chinese outlets' relays of the WeChat article; the report that US federal agencies were told to remove DeepSeek, which we could not trace to a primary document.
• Unconfirmed: any coding benchmark for Super-agent or for Harness as a harness; the model DeepSeek runs internally; V4.1 Pro's date; whether Harness's provider support covers Anthropic and OpenAI natively (reported at launch) or only through custom endpoints; DeepSeek's valuation.