~ / comparisons / china-coding-agents

China's Big-Tech Coding Agents vs Codex and Claude Code

ByteDance Trae, Tencent's CodeBuddy and WorkBuddy, Baidu Comate, Alibaba's Qwen Code and Qoder, Z.ai Zcode and DeepSeek Harness now match Codex and Claude Code on the feature surface. None has been independently measured on DeepSWE or Terminal-Bench, so AgentRanks does not rank them as reaching Codex/Claude Code level. This page separates what is verified from what is vendor marketing.

The reference bar

“Reaching Codex / Claude Code level” has a measured meaning here, not a vibes one.

ReferenceWhat it isScore
Claude CodeAnthropic terminal agent, the AgentRanks architecture leader86
ChatGPT CodexOpenAI terminal agent, the second architecture slot84
Claude Code + Fable 5Terminal-Bench 2.1 stack leader (agent harness + model together)83.8%

A Chinese big-tech agent “reaches Codex / Claude Code level” only once it carries an AgentRanks architecture score near 84–86, or a Terminal-Bench stack score in the same band. Until then, feature parity is a product claim, not a ranking finding.

Star scale

What a star means, and where stars are withheld.

5★ = Opus 5 / Fable 5 measured level (Terminal-Bench frontier) · 1★ = basic. Coding-ability stars are withheld for every entry below: none has been independently measured on DeepSWE or Terminal-Bench, and that gap is the point of this page. Where a third-party editorial score exists (RECATOOLS /10), it is shown on each profile as a source-labelled bar — an editorial rating, not a benchmark.

Feature and benchmark comparison

Feature surface is real; the benchmark column is the gap. Click any name for its full profile.

AgentVendorFormAgent / plan modeMCPARscoreUsersSource
Claude CodeAnthropicCLI/DesktopYesYes86not published
ChatGPT CodexOpenAICLI/DesktopYesYes84not published
TraeByteDanceIDE + CLIBuilder + SOLOYesnot scored6M registered (first year)Official
CodeBuddyTencentIDE + plugin + CLICraft + PlanYesnot scorednot published
WorkBuddyTencentDesktop workplaceSkills + automationYesnot scorednot published
ComateBaiduPlugin + IDEZuluYesnot scored7.6M developers (May 2025)Official
Qwen CodeOSAlibabaCLI45open-source, GitHub-tracked
QoderAlibabaIDE + CLIRepo Wiki + QuestYesnot scorednot published
ZcodeZ.aiDesktop ADEGoal modeYesnot scorednot published
DeepSeek HarnessOSDeepSeekCLI + local web UIPlugin kernelpluginnot scored~169K GitHub starsMarketplace

User figures are vendor-reported unless marked Marketplace or Press, and none is independently audited. “Not scored” means the agent is outside the AgentRanks architecture rubric; it is not a low score.

Use-case capability matrix

Scenario coverage — presence, not benchmark performance.

AgentLocal filesDocs / PPTExcel reportsAnalysisWeb searchMini-programsWeb building
Trae
CodeBuddy
WorkBuddy
Comate
Qwen Code
Qoder
Zcode??
DeepSeek Harness

✓ supported · ◐ partial or via a sibling product (e.g. QoderWork CN, WorkBuddy Skills) · ? unclear · — not in scope. Click an agent name for per-scenario detail and official links.

Per-agent notes

What each entry ships, with its pricing tier and provenance.

AgentWhat it isPricing tierSource
TraeByteDance's VS Code fork IDE with Builder + SOLO; open-source Trae-Agent CLI (MIT).Free tier + $3 / $10 / $30 / $100RECATOOLS / Press
CodeBuddyTencent's coding arm (Buddy AI); first Chinese assistant with MCP; Code 2.0 added sub-agents.Free / ¥99 / ¥199 / ¥999RECATOOLS / Press
WorkBuddyTencent's desktop workplace agent (Buddy AI); local files, PPT, Excel, web automation, Skills.Free / ¥99 / ¥199 / ¥999Vendor (Bench)
ComateBaidu's ERNIE assistant (June 2023); Zulu agent runs requirement-to-code; CAICT 4+ rated.Free / Pro ¥55Official / CAICT
Qwen CodeOSAlibaba's open-source CLI; the only scored Chinese big-tech entry (ARscore 45).Free (open source)AgentRanks
QoderAlibaba's first-party IDE (Aug 2025 preview; CN renamed from Tongyi Lingma 2026-05-20); Repo Wiki + Quest.Pro $20 / $60 / $200RECATOOLS
ZcodeZ.ai's desktop ADE, official harness for GLM-5.2 / 5.3; Goal mode, multi-agent, remote control.Client free; $18 / $72 / $160bitdoze / HN
DeepSeek HarnessOSDeepSeek's MIT agent harness (Aug 13, 2026); Cordis plugin kernel; provider-agnostic.Free (MIT); pay-per-token APIVentureBeat / GitHub

Qwen Code vs Qoder: distinct Alibaba products, not the same tool renamed. Qwen Code is the open-source Qwen-first CLI (scored 45 here); Qoder is Alibaba’s first-party agentic IDE (Repo Wiki + Quest, multi-model routing, not scored). CodeBuddy vs WorkBuddy: Tencent’s Buddy AI splits coding (CodeBuddy, IDE+CLI) from workplace (WorkBuddy, desktop files/PPT/Excel/automation); the two share one credit account.

The honest verdict

Feature parity is real; benchmark parity is unproven.

The eight Chinese big-tech entries ship the same agentic surface that defines Codex and Claude Code in 2026: multi-file generation, plan-first execution, MCP for external tools, and a CLI or IDE path. On features, they are in the conversation. On benchmarks, they are not: none publishes or has been independently measured on DeepSWE or Terminal-Bench 2.1, so AgentRanks cannot place any of them next to Codex (84) or Claude Code (86). “Reaching Codex / Claude Code level” is, for now, a vendor-marketing claim rather than a ranking finding — it becomes a finding the day one of them is scored.

Compare further

Reference agents and the scored Chinese entry.

Claude Code

the architecture leader

score 86 / CLI/Desktop

ChatGPT Codex

the reference OpenAI terminal agent

score 84 / CLI/Desktop

Qwen CodeOS

the only scored Chinese big-tech entry

score 45 / CLI

Claude Code vs Codex

the two reference agents head to head

long-tail compare