Foundation Lab Coding Agents

Executive Summary

Every major AI foundation lab now ships its own official coding agent. What was a four-horse Western race in March is now a seven-entrant global category: Anthropic, OpenAI, Google, and xAI were joined by Alibaba, Moonshot AI, and Mistral. These foundation lab coding agents are tools built by the model makers themselves, with native optimizations and integrations unavailable to third-party alternatives.

Products: Claude Code (Anthropic), Codex (OpenAI), Gemini CLI and Antigravity CLI (Google), Grok Build (xAI), Qwen Code (Alibaba), Kimi Code (Moonshot), Mistral Vibe (Mistral)

What changed since March:

• Google blinked on open source — individual Gemini CLI tiers sunset June 18, 2026, replaced by the closed-source, Go-based Antigravity CLI; enterprise Code Assist licenses retain Gemini CLI^[1][2]
• Codex became a business — $1B+ annualized revenue, 4M+ weekly developers, GPT-5.5 default, Windows app, Chrome extension, mobile^[3][4]
• Grok Build actually shipped (May 14, early beta) — a Rust CLI with Plan Mode and worktree-isolated parallel subagents; the pre-launch rumors this report previously carried (8 agents, Arena Mode, local-first) did not survive contact with the real product^[5]
• The Chinese labs arrived — Qwen Code (25K stars) and Kimi Code (K2.6-powered) are GA-grade tools, monetizing through coding plans just like the Western labs^[6][7]
• Prices went up everywhere — Claude Code left the $20 Pro plan; Codex restructured; Qwen killed free OAuth^[8]

Strategic Planning Assumptions:

• By 2027, every foundation lab agent will have parallel/background execution (five of seven already do)
• By 2028, foundation lab agents will capture more than 50% of coding agent market share
• DeepSeek is hiring a "Code Harness" team (May 2026) — expect an eighth entrant within two quarters

---

Market Definition

Foundation lab coding agents are AI coding tools built and maintained by the companies that train the underlying foundation models.

Inclusion Criteria:

• Built by the foundation model provider (not a third-party wrapper)
• Designed primarily for code generation, editing, and developer workflows
• Available as a downloadable/installable tool (CLI, app, or extension)
• General availability or public preview

Exclusion Criteria:

• Third-party tools using foundation model APIs (Cursor, Windsurf, etc.)
• Chat interfaces without agentic capabilities (ChatGPT web, Claude.ai)
• IDE plugins not maintained by the model provider
• Platform vendors that don't train the frontier models they serve (GitHub Copilot CLI, Amazon Q Developer)
• Model subscriptions without a first-party tool (Z.ai's GLM Coding Plan; DeepSeek — nothing shipped yet)

Why This Category Matters: Foundation lab agents have unfair advantages: first access to new model capabilities, deeper integration with safety systems, and optimization feedback loops that improve both the agent and the model.

---

Comparison Matrix

---

Product Profiles

Claude Code

The capability and revenue benchmark^[9][10]

• $2.5B+ run-rate revenue; enterprise is over half; 131K GitHub stars
• Surfaces: terminal, IDE, web, mobile, Slack; background sessions, scheduled Routines, subagent orchestration, computer use
• Current models: Fable 5, Opus 4.8 (default on Max/Team Premium), Sonnet 4.6
• Corrected from the March edition: the repo is public but the license is proprietary — Claude Code is not open source; and it left the $20 Pro plan in April ($20–125/seat across tiers)

---

Codex

The adoption leader^[4][11][3]

• 4M+ weekly developers, ~$1B+ annualized revenue, 90K stars
• Surfaces: macOS and Windows apps, CLI, IDE extensions, web/cloud, Chrome extension, mobile, Slack/Linear/GitHub
• GPT-5.5 frontier default (the codex-1/o3 era is over); built-in review agent; leads Terminal-Bench 2.0
• Pricing restructured: Go $8/mo, Plus $20, Pro from $100 with 5x/20x tiers

---

Gemini CLI

The open-source champion, sunsetting^[2][1]

• 105K stars — the most-starred tool in the category, and Google is walking away from its individual tiers on June 18, 2026
• Successor: Antigravity CLI — closed-source, Go-based, multi-agent, weekly quotas; community reaction has been hostile
• Enterprise continuity: Gemini Code Assist Standard ($19/user/mo annual) and Enterprise ($45) retain Gemini CLI
• Until sunset: 1,000 free requests/day, Gemini 3 family, 1M context

---

Antigravity CLI

The closed-source successor^[1]

• Go-based agy binary running the multi-agent Antigravity 2.0 harness with Gemini 3 Pro on all tiers; the I/O demo showed 93 subagents building an OS core in ~12 hours
• Free tier with weekly quotas; AI Pro/Ultra quotas refresh every five hours after the rate-limit backlash
• Closed source with undisclosed internals — the trade Google made for Gemini CLI's 105K-star openness; community reaction has been hostile

---

Qwen Code

Alibaba's GA-grade open CLI^[6]

• 25K stars, Apache 2.0, 481 releases (v0.17); headless/CI mode, IDE plugins, SDKs, daemon mode
• Qwen3.6-Plus default but genuinely multi-provider (OpenAI, Anthropic, Gemini, local via Ollama/vLLM)
• Free OAuth tier ended April 15, 2026 — API pay-per-use or Alibaba Cloud Coding Plan; self-hosted remains free

---

Kimi Code

Moonshot's K2.6-powered terminal agent^[7]

• TypeScript rewrite (June 2026) of the October 2025 kimi-cli; MIT; ACP for Zed/JetBrains, VS Code extension, video input
• Kimi K2.6: 1T-parameter MoE, 256K context, strong SWE-bench Pro claims (vendor-reported)
• ~$19/mo coding membership or pay-as-you-go API; the rewrite is weeks old — expect churn

---

Mistral Vibe

Europe's entry — minimal, open, remote-capable^[12]

• 4.4K stars, Apache 2.0; launched December 2025 with Devstral 2, now powered by Mistral Medium 3.5
• Remote cloud agents launchable from the CLI or Le Chat; Zed/ACP integration
• The lightest-weight tool in the category by design

---

Grok Build

Shipped, finally — smaller than the rumors^[5]

• Early beta since May 14, 2026: Rust CLI with Plan Mode, parallel subagents in isolated git worktrees, headless CI mode, MCP/AGENTS.md/plugins
• Powered by grok-build-0.1 (256K context); no published benchmarks for the shipped model
• API $1/$2 per M tokens; launched on SuperGrok Heavy ($300/mo), expanded to SuperGrok and X Premium Plus
• The March edition's "8 parallel agents, Arena Mode, local-first" claims came from pre-launch rumor coverage and were wrong — corrected throughout

---

Gap Analysis

Gap insights:

• Qwen Code is the only multi-provider lab tool — an open question whether the Western labs ever follow
• Issue-tracker depth remains thin — Codex has Linear; Jira is absent everywhere; the orchestration gap above individual agents persists
• The open-source center of gravity shifted east and to Europe — with Gemini CLI sunsetting, the OSI-licensed options are Qwen, Kimi, and Mistral

---

Strategic Recommendations

By Use Case

By Buyer Profile

Individual developers: the free-tier era is ending — Qwen Code self-hosted or Mistral Vibe with pay-per-use API is the new budget path.

Claude-committed teams: Claude Code, now Team Premium-gated ($100–125/seat for Claude Code access) — budget accordingly.

OpenAI ecosystem: Codex's unified account across Mac/Windows/CLI/web/mobile is the most complete experience in the category.

Enterprises needing model flexibility, Jira, signed commits: foundation lab agents still don't solve this. Tools like Tembo fill the orchestration gap by sitting above individual agents.

---

Market Outlook

Near-Term (2026)

• The Gemini CLI sunset (June 18) sends individual users to Antigravity, Qwen Code, or out of the Google ecosystem entirely — watch the star-migration
• DeepSeek's Code Harness team ships an eighth entrant
• Grok Build either reaches GA with benchmarks or confirms the skeptics

Medium-Term (2027-2028)

• Enterprise integrations (Jira, SSO, signed commits) arrive lab by lab
• Pricing stratifies further: capability tools ($100+/seat) vs open CLIs (free + API)
• Model-native optimization creates measurable productivity gaps vs third-party tools

Long-Term (2029+)

• Foundation lab agents capture majority market share
• Third-party tools survive in orchestration, compliance, and multi-model niches
• Agent-to-agent interoperability (MCP, ACP, A2A) determines whether lock-in holds

---

Bottom Line

The category doubled and stratified in three months. The economics are now visible: Codex at $1B+ ARR and Claude Code at $2.5B+ run-rate prove first-party agents are core lab businesses — which is exactly why the free and open era is receding (Gemini CLI sunset, Qwen OAuth shutdown, the price-hike wave).

Current positioning (June 2026):

• Claude Code — capability and revenue leader; no longer cheap, never open
• Codex — adoption and surface leader; the most complete consumer-to-enterprise spread
• Qwen Code — the open-source standard-bearer now that Gemini CLI is sunsetting
• Kimi Code / Mistral Vibe — credible open entrants from Moonshot and Mistral
• Gemini CLI — 105K stars and a eulogy date
• Grok Build — real but early; judge it again next quarter

The trade-off is still lock-in — only Qwen Code breaks the pattern. For enterprises needing model flexibility and workflow integrations, the orchestration layer above these agents remains the open opportunity.

---

Research by Ry Walker Research • methodology

Disclosure: Ry Walker is CEO of Tembo, which offers AI coding agent orchestration.