Skip to content
Capital & Compute
· ai· coding-agents· tools

New Agentic Code Editors in 2026: ZCode, Antigravity 2.0

ZCode, Antigravity 2.0, Grok Build, Warp, and Zed all changed in 2026. Here is what actually shipped in each new agentic code editor, and what it costs.

By Capital & Compute

Four genuinely new agentic code editors shipped between mid-May and the end of June 2026: Google’s Antigravity 2.0, xAI’s Grok Build, Z.ai’s ZCode, and Zed’s Terminal Threads. One older tool made a very 2026 pivot in the same window: Warp open-sourced its client. ByteDance’s Trae gets a mention too, not because it changed, but because it’s the one tool here that’s still free with no catch. The site’s AI coding agents hub tracks the whole field; this is what changed in it over the last two months, and what each change actually costs to use.

  • ZCode (Z.ai, June 2026): official desktop harness for GLM-5.2, built around Goal Mode and sub-agents.
  • Google Antigravity 2.0 (May 19, 2026): replaces Gemini CLI entirely by June 18, 2026.
  • Grok Build (xAI, May 14, 2026, beta): terminal-native agent with subagents in isolated Git worktrees.
  • Zed Terminal Threads (May 20, 2026): runs Claude Code, Amp, or Codex as a managed thread in Zed’s sidebar.
  • Warp (open-sourced April 28, 2026): the existing Agentic Development Environment’s client is now AGPLv3.
  • Trae (ByteDance): not new, but still fully free with no paid tier, unlike everything else on this list.

None of these six replace Claude Code, Cursor, or Copilot as the default choice for most developers. What five of them do is worth tracking anyway, because the pattern is the same across all of them: every major coding-agent maker is racing to ship subagents, background tasks, and terminal-native control in the same eight-week window. That is not a coincidence, and it changes what “agentic editor” means going into the second half of 2026.

Why five tools moved at once

Google, xAI, Z.ai, Warp, and Zed are not copying each other by accident. Each is reacting to the same pressure: a single AI agent that edits one file at a time is no longer the state of the art, and hasn’t been since parallel subagents (isolated instances that can work on separate parts of a codebase at once, then report back) became the feature customers ask about first. Antigravity 2.0 frames it as agents “communicating with each other.” Grok Build frames the same idea as subagents running in isolated Git worktrees. Different vocabulary, same bet: one agent per task, running at once, instead of one agent doing tasks in sequence.

The other shared thread is distribution. Google is folding Antigravity into the same account system as Gemini. xAI is bundling Grok Build into SuperGrok and X Premium+, subscriptions people already pay for something else. Z.ai built ZCode specifically to sell more GLM Coding Plan seats. None of these five shipped as a standalone business; all five are a wrapper around a subscription the maker already wants you on.

ZCode: Z.ai’s official harness for GLM-5.2

ZCode is Z.ai’s (the company behind the GLM model family, formerly branded Zhipu AI) desktop agent built specifically around GLM-5.2. Z.ai’s own ZCode documentation describes it as an “Agentic Development Environment” rather than a code-completion plugin: instead of one instruction at a time, you define a “Goal,” and ZCode plans, executes, and verifies the work across multiple steps before reporting back.

The feature set beyond Goal Mode: MCP server support, sub-agents for parallel work, a desktop app for macOS, Windows, and Linux, and remote control from mobile or from Feishu/WeChat bot channels, useful if you want to kick off a long task from a phone and check on it later. ZCode does not run its own model. It runs on GLM-5.2 (with GLM-5-turbo as a lighter alternative), the same model this site covered when GLM-5 shipped from vibe coding to agentic engineering. Z.ai’s own ZCode documentation advertises up to 1 million tokens of context for long-horizon tasks, roughly four times the context window xAI states for Grok Build below.

Access runs through the GLM Coding Plan rather than a separate ZCode subscription. The verified entry-level price on this site’s AI pricing tracker is the GLM Coding Plan Lite tier at $18/month (verified June 14, 2026, against Z.ai’s own subscribe page); Z.ai has not published a separate, verified dollar figure for the Pro and Max tiers ZCode’s own docs allude to, so treat any specific number above $18 you see quoted elsewhere as unverified until Z.ai’s pricing page states it directly. Through July 31, 2026, Z.ai is running a discounted-quota promotion for GLM Coding Plan subscribers and a 5-day free trial for new users, per the ZCode docs above.

Google Antigravity 2.0: Gemini CLI’s replacement, not a side project

Google announced Antigravity 2.0 on May 19, 2026, at Google I/O, and it is not an incremental update. According to Google’s own developer blog, Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for free and paid consumer accounts on June 18, 2026. If you were running Gemini CLI, Google’s message is not “here’s an upgrade,” it’s “here’s the thing that replaces what you were using, and the old one turns off in a month.” Enterprise customers keep access through paid API keys; everyone else is migrated by default.

What replaces it is the Antigravity CLI, which Google’s blog confirms retains Agent Skills, Hooks, Subagents, and Extensions from Gemini CLI (Extensions are now called Antigravity plugins) rather than dropping that surface area. Beyond the CLI, the Antigravity 2.0 launch bundled a standalone desktop app, an Antigravity SDK for building custom agents on the same harness Google uses internally, and Gemini 3.5 Flash as the new default model, according to TechCrunch’s coverage of the I/O 2026 keynote. This site already covered Gemini 3.5 Flash’s cost per coding task on its own; Antigravity 2.0 is the harness Google now wants that model running inside.

This is the deep-dive this site’s AI coding agents hub had flagged as a gap. Antigravity had a row in the hub’s comparison table with no dedicated coverage behind it; this is that coverage.

Grok Build: xAI’s terminal agent, and a benchmark claim worth being careful with

xAI shipped Grok Build into beta on May 14, 2026: a terminal-native coding agent built on the grok-build-0.1 model, written in Rust, with a 256,000-token context window. Plan Mode is on by default. Before Grok Build touches a file, it proposes a step-by-step plan you can approve, comment on individual steps of, or rewrite outright; nothing executes until you sign off. For larger tasks, it delegates to subagents that run in parallel, each in its own isolated Git worktree, so two agents editing different parts of a codebase can’t collide with each other’s uncommitted changes. Access runs through SuperGrok and X Premium+ subscriptions rather than a separate Grok Build plan; xAI’s own pricing page has the current subscription figures, which are not yet in this site’s verified pricing dataset.

Warp: not a new launch, but a real pivot

Warp 2.0 itself is not 2026 news: Warp shipped it as “the first Agentic Development Environment” back in mid-2025. What happened in 2026 is that Warp open-sourced its client entirely. On April 28, 2026, Warp announced that its codebase is now available on GitHub under the AGPLv3 license, with OpenAI as the founding sponsor of the new open-source repository. Warp is calling the framing “Open Agentic Development”: Oz, Warp’s own cloud agent orchestrator, handles most of the implementation work on community-approved GitHub issues, while the community shapes direction and verifies the results, rather than writing most of the code by hand. The same release added routing to open-weight models (Kimi, MiniMax, Qwen) through a new “auto (open)” option that picks the best open model for a given task automatically.

Why this matters for a “new editors” roundup: an open-sourced, AGPLv3-licensed terminal that can spawn Claude Code, Codex, or Antigravity CLI as managed sub-agents is a meaningfully different competitive object than the same terminal was in 2025, even though the underlying product name didn’t change.

Zed’s Terminal Threads: a direct response to Anthropic’s pricing change

Zed shipped Terminal Threads on May 20, 2026, in version 1.3.5. The feature, announced on Zed’s blog, lets you run claude, amp, codex, pi, or any other terminal-based process as a managed thread inside Zed’s Agent Panel sidebar, tracked and revisited the same way as Zed’s native agent threads.

The blog post is unusually direct about why Zed built this now: Anthropic is moving Agent SDK usage on subscription plans to a separate, limited credit system starting June 15, 2026, which Zed says makes running Claude Code through the Agent Client Protocol (ACP) 15 to 30 times more expensive for heavy usage. Terminal Threads sidestep that entirely, because a terminal-run claude process bills against your existing Claude subscription the normal way claude always has, not against the new ACP credit system. Zed’s own framing: Terminal Threads are “the only way to keep using Claude Code in Zed with your existing Claude subscription” after that change lands. This is the same subscription-versus-metered-usage split this site has covered in what Claude Code actually costs per task; Zed built a whole feature specifically to keep its users on the cheaper side of that line.

Trae: the counterexample, not a new launch

Trae, ByteDance’s VS Code fork, isn’t news this cycle, and unlike the rest of this list it hasn’t added a paid tier either. Trae’s own site states plainly that it’s free with no hidden costs, full stop, not a free tier bolted onto a paid product. Builder Mode, the feature that reads a request, breaks it into steps, and shows a live preview before executing, ships on that same free plan. So does MCP support, including third-party integrations like Figma.

Trae earns a mention here for one reason: it’s the only tool in this piece that’s still free by default rather than free-with-a-catch. Everything else on this list (ZCode, Antigravity, Grok Build, Warp, Zed’s paid-adjacent Terminal Threads workaround) assumes you’re already paying for something.

Context window size: ZCode vs. Grok BuildBar chart comparing stated context window size in tokens. ZCode, running on GLM-5.2, is stated at 1,000,000 tokens. Grok Build, running on grok-build-0.1, is stated at 256,000 tokens.0K tokens200K tokens400K tokens600K tokens800K tokens1,000K tokensZCode (GLM-5.2)1,000K tokensGrok Build (grok-build-0.1)256K tokens
Context window size: ZCode vs. Grok Build
ItemValue
ZCode (GLM-5.2)1,000K tokens
Grok Build (grok-build-0.1)256K tokens
Context window size, in tokens, for the two new editors whose makers published a specific figure. ZCode runs on GLM-5.2, which Z.ai states can hold up to 1 million tokens for long-horizon tasks. Grok Build runs on grok-build-0.1, which xAI states carries a 256,000-token window.Source: Z.ai, ZCode Docs (2026); xAI, Introducing Grok Build (2026)

The other four tools in this piece (Antigravity 2.0, Warp, Zed Terminal Threads, Trae) don’t publish a comparable single context-window figure for the harness itself, since Antigravity and Zed both let you swap the underlying model, and Warp and Trae are model-agnostic by design.

The six tools, side by side

Tool Maker What changed Interface Pricing model
ZCode Z.ai New desktop harness for GLM-5.2 Desktop app, mobile remote, bot channel GLM Coding Plan (from $18/mo, verified)
Antigravity 2.0 Google Replaces Gemini CLI entirely by June 18, 2026 Desktop app, CLI, SDK Bundled with Google AI plans
Grok Build xAI New terminal agent, beta Terminal CLI SuperGrok / X Premium+ subscription
Warp Warp Client open-sourced (AGPLv3) Terminal Free client; model usage billed separately
Zed Terminal Threads Zed Lets other agents run inside Zed’s sidebar Editor sidebar Free (Zed); billed by whichever agent you run
Trae ByteDance No change; included as the free counterexample Editor (VS Code fork) Free, no paid tier

For the tools this site already covers in depth, cost is the deciding factor more often than features: see Cursor Composer 2.5’s real cost, what Windsurf (now Devin Desktop) actually costs, and Kiro’s pricing explained. The 2026 AI coding agent landscape is the broader survey this piece updates.

Bottom line

If you already pay for Claude Code, Cursor, or Copilot, none of these six require you to switch. What they signal is where the money and engineering effort in this category is actually going next: parallel subagents, terminal-native control, and tighter bundling into subscriptions you already have. Antigravity 2.0 already forced a decision for Gemini CLI users: free and paid consumer access stopped on June 18, 2026, so if you were still on Gemini CLI by then, you were migrated to the Antigravity CLI whether you planned for it or not. Zed’s Terminal Threads is the one with an ongoing dollar impact if you run Claude Code inside an ACP-connected editor, since Anthropic’s credit-system change on June 15, 2026 is already in effect. The rest are worth knowing about, not worth switching for yet.

Frequently asked questions

What are the newest agentic code editors in 2026?
Four editors are genuinely new since mid-May 2026: Google Antigravity 2.0 (May 19), xAI Grok Build (May 14, beta), Z.ai ZCode (June), and Zed Terminal Threads (May 20). Warp is an older product that made a major change in the same window, open-sourcing its client on April 28, 2026. Trae, ByteDance's editor, did not change; it is included as the one tool here that remains fully free with no paid tier.
Is ZCode free?
ZCode itself has no separate price. Access runs through the GLM Coding Plan, whose verified entry-level Lite tier costs $18/month as of June 14, 2026. Z.ai has not published verified dollar figures for higher tiers. New users get a 5-day free trial, and a discounted-quota promotion runs through July 31, 2026.
What happens to Gemini CLI users after Antigravity 2.0?
Gemini CLI and the Gemini Code Assist IDE extensions stop serving requests for free and paid consumer accounts on June 18, 2026, per Google’s own developer blog. Enterprise customers keep access through paid API keys. Everyone else needs to move to the Antigravity CLI, which keeps the same Agent Skills, Hooks, Subagents, and Extensions (now called plugins) that Gemini CLI had.
Does Grok Build really score 70.8% on SWE-bench Verified?
That figure is widely repeated but traces back to grok-code-fast-1, an earlier xAI model, not to grok-build-0.1, the model actually running inside Grok Build. xAI has not published an independently verified SWE-bench Verified score for the production Grok Build agent.
Why did Zed build Terminal Threads?
Zed says it built Terminal Threads because Anthropic is moving Agent SDK usage on subscription plans to a separate, limited credit system starting June 15, 2026, which makes running Claude Code through Zed’s Agent Client Protocol integration 15 to 30 times more expensive for heavy usage. Running Claude Code as a Terminal Thread instead bills against your normal Claude subscription, not the new credit system.

Sources

Subscribe to Capital & Compute

Source-backed analysis of what AI compute really costs, sent when a new post goes live.

No spam. Unsubscribe anytime.

← Back to all posts