Coding Agents in 2026: Codex vs Claude Code vs Antigravity vs Copilot

By mid-2026, coding agents have moved from experimental novelty to the default way professional developers build software. Four platforms dominate the conversation: OpenAI’s Codex , Anthropic’s Claude Code , Google’s Antigravity , and the incumbent GitHub Copilot . Each takes a fundamentally different approach to the same problem — how much of the software development lifecycle can an AI own, and how much control stays with the human.

This is a practical guide to the current state: what each tool does well, where it falls short, and which stack makes sense for your team.

OpenAI Codex: From CLI to Superapp

OpenAI released Codex as a CLI tool in April 2025 and followed with desktop apps for Windows and macOS in February 2026. The Windows version has native sandboxing — it runs agent actions under restricted tokens with file-system ACLs, so an autonomous coding session can’t accidentally modify system files. This makes it the most security-conscious offering for Windows shops.

Models and Variants

Codex is powered by GPT-5.4 (upgraded from GPT-5.3-Codex in March 2026). OpenAI also ships Codex-Spark, a variant running on Cerebras hardware that delivers about 15x speedup for batch code generation tasks. For teams that process large codebases in CI pipelines, Spark changes the economics of AI-driven refactoring.

Codex Security

In March 2026, OpenAI launched Codex Security — a vulnerability scanning layer that sits on top of the coding agent. It has already scanned 1.2 million commits and surfaced 800 critical vulnerabilities. The tool integrates directly with GitHub Agent HQ, so findings appear alongside pull requests without extra tooling.

Enterprise Traction

Codex reached 2 million weekly active users by March 2026, up from about 5% of Claude Code’s usage in September 2025 to about 40% by January 2026. Enterprise customers include Cisco, Nvidia, Ramp, Rakuten, and Harvey. OpenAI is also merging Codex into a desktop superapp alongside ChatGPT, which should further accelerate adoption among teams already in the OpenAI ecosystem.

Integrations

Codex ships with plugins for enterprise workflows, plus direct integrations with Figma (for generating UI from design files) and Xcode (for iOS development). A May 2026 Dell partnership brings Codex to hybrid and on-premises enterprise environments through the Dell AI Data Platform and Dell AI Factory.

Anthropic Claude Code: The CLI Powerhouse

Claude Code launched as a command-line agent in February 2025 and reached general availability in May 2025. Its terminal-native design appealed to developers who prefer keyboard-driven workflows. A web version with sandboxing followed in October 2025.

Claude Opus 4.7 and Beyond

The current model powering Claude Code is Claude Opus 4.7, released in April 2026. Anthropic has continued pushing on code generation quality — Claude Code found over 100 bugs in the Firefox codebase, including 14 high-severity issues. In a benchmark project with Claude Opus 4.6, 16 agents working in parallel wrote a C compiler in Rust from scratch, the first model capable of the feat (even if the resulting compiler isn’t very efficient).

Dreaming: Persistent Memory

The May 2026 Dreaming feature gives Claude Code persistent memory consolidation across sessions. After completing a task, the agent can reflect on the codebase structure, store learned patterns, and apply them to future work. This makes it well suited for long-running maintenance projects where context builds over weeks.

Enterprise Adoption

Claude Code’s revenue jumped 5.5x by July 2025. Notable users include NASA, which used Claude Code to plan Mars rover routes, and Norway’s $2.2 trillion sovereign wealth fund. Microsoft and Google employees reportedly use Claude Code. OpenAI employees used it too until Anthropic revoked their access in August 2025 for violating terms of service. Claude experienced 4.9% month-over-month subscription growth in February 2026 — contrasting with OpenAI’s -1.5% in the same period.

The Broader Claude Ecosystem

Anthropic has expanded beyond the CLI. Claude Cowork (January 2026) provides a GUI version for less technical team members, and Claude Design (April 2026) adds visual creation capabilities. Claude Code Security, released in February 2026, reviews entire codebases for vulnerabilities — competing directly with Codex Security.

Google Antigravity: Agent-First IDE

Announced on November 18, 2025 alongside Gemini 3, Google’s Antigravity is Google’s most ambitious rethinking of the developer environment. It’s an AI-first IDE — forked from VS Code (or possibly Windsurf, depending on who you ask) — but the comparison stops there.

Powered by Gemini 3.1 Pro and Gemini 3 Flash, Antigravity also supports third-party models including Claude Sonnet 4.6, Claude Opus 4.6, and the open-source GPT-OSS-120B. The latest release, version 2.0.1, came out on May 19, 2026.

Manager View: Parallel Agent Orchestration

Antigravity’s defining feature is the Manager view, which lets you orchestrate multiple parallel agents across separate workspaces. One agent can refactor an API while another writes integration tests and a third reviews the documentation — all running concurrently in isolated sandboxes. Agents produce Artifacts that include task lists, implementation plans, screenshots, and browser recordings.

Editor View and Pricing

The Editor view offers a more traditional IDE experience with an agent sidebar, integrated browser for agent use, and full terminal access. Antigravity is free during its public preview with generous rate limits.

Limitations and Risks

Antigravity is the newest entrant with the smallest user base and no public adoption metrics. It’s still in preview — pricing after the preview ends has not been announced, creating uncertainty for teams considering a commitment. The tool is deeply tied to Google’s ecosystem: while it supports third-party models, the best experience requires Gemini. There’s also the unresolved question: is it a straight VS Code fork or a Windsurf fork? The answer matters for plugin compatibility and community support. For teams already on Google Cloud or using Gemini extensively, these are manageable risks. For everyone else, Antigravity is best evaluated on a side project before any production commitment.

GitHub Copilot: The Incumbent Strikes Back

GitHub Copilot remains the most widely adopted AI development tool by raw user count. Launched in 2021, it spent its first four years as an intelligent autocomplete and only gained true agent capabilities in 2025.

Agent mode arrived in February 2025, giving Copilot autonomous editing within VS Code. The async coding agent followed in May 2025 — it can create pull requests without blocking the developer. Between May and September 2025, GitHub reported over 1 million PRs created by AI agents on the platform.

Multi-Model Support

Unlike the other platforms tied to their own models, Copilot now supports Anthropic, Google, and OpenAI models. Pro+ subscribers can even delegate tasks to third-party agents like Claude Code and Codex directly from the Copilot interface.

Pricing

Copilot’s pricing remains its strongest advantage:

Business and Enterprise plans include IP indemnity, which is a deciding factor for regulated industries. Copilot supports every major IDE: VS Code, Visual Studio, JetBrains, Xcode, Neovim, and Eclipse.

Head-to-Head Comparison

Note: Model Context Protocol (MCP) support is included where noted.

Adoption Metrics and What They Mean

The numbers speak for themselves. Claude Code’s 5.5x revenue jump and 4.9% monthly subscription growth point to strong enterprise momentum. Codex’s user growth — from 5% to 40% of Claude Code’s volume in four months — shows OpenAI’s distribution advantage with the ChatGPT user base.

Copilot leads on total PRs created (1 million in five months), but that figure includes simpler autofixes alongside full agent-driven changes. Still, it’s the clearest signal that teams are trusting AI agents with production code. Antigravity remains the newest entrant with the smallest user base and no public adoption data, yet its Manager view represents a different paradigm — one that may scale better for large codebases.

SWE-bench Perspective

Standardized benchmarks offer another view. Claude models have consistently led SWE-bench Verified scores since Claude 4’s release, especially on complex multi-file editing tasks. Codex has closed the gap with GPT-5.4, particularly on Python and TypeScript benchmarks. Copilot’s scores depend on the underlying model selected — using Claude Opus 4.7 through Copilot yields similar results to running Claude Code directly. Antigravity’s Gemini 3.1 Pro performs competitively on agentic coding benchmarks but has less third-party validation than the others. Benchmark scores only tell part of the story — pair them with real-world workflow fit.

Decision Matrix: Which Stack for Which Team?

Solo developer or small startup

Choose Copilot Pro ($10/mo) for breadth of IDE support and lowest cost. If you prefer terminal workflows, Claude Code with the Claude Pro subscription gives you Opus 4.7 quality and the Dreaming memory feature for long-running projects.

Mid-size team (5-50 engineers)

Codex becomes attractive here, especially if you’re on Windows or use Figma/Xcode. The enterprise plugin system and Codex Security layer reduce the burden on your security team. Compare against Claude Code — the 5.5x revenue growth signals strong enterprise momentum, and the Dreaming feature helps teams that maintain large monorepos.

Large enterprise with compliance requirements

Copilot Business/Enterprise ($19-$39/mo per seat) offers IP indemnity, the widest IDE support, and integration with existing GitHub workflows. Pair it with Codex Security or Claude Code Security for vulnerability scanning. The multi-model support in Copilot means you aren’t locked into a single AI provider.

Teams building agent-driven pipelines

Antigravity’s Manager view is unique — no other tool lets you run parallel agents across workspaces with built-in orchestration. If you’re willing to bet on Google’s infrastructure and can live with the preview-stage roughness, this is the most forward-looking option. Its free pricing during preview makes it low-risk to evaluate.

The Takeaway

No single coding agent wins for every scenario. Copilot has the distribution and pricing. Claude Code has the deepest model and the most loyal users. Codex has the security story and the Windows sandboxing. Antigravity has the boldest vision for what an IDE can become.

The smartest strategy in 2026 is to stay multi-model. Here’s a concrete 6-month plan:

1. Keep Copilot Pro ($10/mo) as your daily driver for IDE integration across all your editors
2. Run Claude Code for complex refactoring — the Dreaming memory feature pays for itself on multi-week projects where context matters
3. Spin up Antigravity on a side project to evaluate the Manager view with no financial commitment (it’s still free)
4. Use Codex if you’re on Windows — the native sandboxing is best-in-class for security-conscious teams

All four platforms are iterating faster than ever. The leaderboard will look different in Q4 2026. The best investment you can make today is building team familiarity with multiple agents, so you can pivot when the next breakthrough arrives.

Practical Evaluation Checklist

Beyond feature lists, here’s what to evaluate when choosing a coding agent:

• Data privacy: Where does code get processed? Does the tool offer on-prem or VPC deployment? (Only Codex through Dell and Copilot Enterprise offer on-prem options)
• Cloud dependency: All four tools require persistent internet connectivity — none work fully offline
• Migration path: Switching agents means retraining CI/CD integrations, MCP server configurations, and team workflows. Plan for a 2-4 week transition
• Code ownership: Review each platform’s terms on training data usage — Copilot Business/Enterprise and Claude Enterprise offer data not used for training. Codex and Antigravity’s individual plans may use interactions for model improvement

Further Reading

• OpenAI Codex — Official product page with feature details, enterprise plans, and the Codex Security offering
• Claude Code Documentation — Anthropic’s comprehensive docs covering setup, workflows, Dreaming, and Claude Code Security
• Google Antigravity — Google’s agent-first IDE with the Manager view for parallel agent orchestration
• GitHub Copilot — Plans, pricing, and multi-model support for the most widely adopted AI coding tool
• SWE-bench — Standardized benchmark leaderboard for evaluating coding agent performance across verified software engineering tasks