The AI Coding Agent Field Guide — Codex, Claude Code, Gemini, Copilot & the Open-Source Wave (July 2026)
What changed this week
If you only track three headlines from July 2026, track these:
The big six, side by side
| Agent | Model (default) | Terminal-Bench 2.1 | Superpower | Watch Out For |
|---|---|---|---|---|
| Codex CLI | GPT-5.5 / 5.6 Sol | 83.4% (#1) | Cloud sandboxing, unified CLI + IDE + cloud experience | Tied to the OpenAI ecosystem |
| Claude Code | Claude Opus 4.8 | 78.9% | Reasoning depth (69.2% SWE-bench Pro), 1M context, hooks, Routines, ~5.5x fewer tokens than IDE agents | No free tier — needs Pro/Max or API billing |
| Gemini CLI / Antigravity | Gemini 3.1 Pro | 70.7% | Massive raw context window; free for individuals via Antigravity | Mid-transition; individual free tier on classic CLI ended June 18 |
| GitHub Copilot | Multi-model picker | — | Deepest editor integration; agent mode + cloud agent + CLI in one subscription | Most power hidden behind non-obvious features |
| Grok Build | Grok 4.5 | — | Speed and price ($2/$6 per 1M tokens); co-designed with Cursor | Ten days old — patterns still forming |
| opencode | Any (BYOK) | — | Most-starred open-source agent (~180k ★, MIT); model-agnostic, free | You pay per-token for whichever model you plug in |
One number worth internalizing: token efficiency compounds. Claude Code using roughly 5.5x fewer tokens than a typical IDE agent means the real monthly cost difference is much bigger than the sticker price suggests — and the same task finishes with less context churn.
OpenAI Codex — the benchmark ceiling
Codex is the volume leader and the current benchmark leader. Its bet is cloud sandboxing — every task can run in an isolated environment, locally or async in the cloud, and come back as a pull request. If you already pay for ChatGPT, Codex is effectively included, which makes it the lowest-friction entry point of the big six. It is also accessible programmatically via the API, so agentic coding drops into CI/CD pipelines cleanly.
npm install -g @openai/codex
# Sign in with your ChatGPT account, then:
codex "add rate limiting to the /api/v2/search endpoint, follow existing middleware patterns"Claude Code — reasoning depth and token thrift
Claude Code is the tool engineers mention most when the task is complex multi-file refactoring or architectural work. Three capabilities set it apart in mid-2026: a 1M-token context with unusually good context efficiency (it loads less but uses it better), a hook system for event-triggered automation, and Routines — scheduled or GitHub-triggered runs, effectively cron for your coding agent. It is also the origin of the SKILL.md convention that nearly every other agent now reads.
npm install -g @anthropic-ai/claude-code
# In your repo:
claude
> think hard, then refactor src/pipeline/ to remove the circular
dependency between loaders and validators. Run tests after each file.Gemini CLI → Antigravity — the whole-monorepo play
Gemini CLI’s identity has always been context volume — when you need to hold an entire monorepo in the window at once, it wins. The story right now is the transition: the classic CLI’s individual free tier (the famous 1,000 requests/day) ended June 18, and the product is folding into Antigravity, Google’s parallel-agent platform, which stays free for individuals. If you were on the free CLI, plan the migration now rather than in a crunch.
GitHub Copilot — the one already in your editor
Copilot’s advantage is surface area on one subscription: inline completions with Next Edit Suggestions, agent mode for multi-file work, a cloud agent that turns GitHub issues into PRs, and a CLI with its own plugin marketplace — plus a model picker that includes Claude and GPT models, so you’re not locked to one brain. Most developers still use a tenth of it. I wrote the full deep-dive on the hidden 90% — custom instructions, prompt files, agents, MCP, /chronicle, Autopilot — in my GitHub Copilot Hidden Tricks guide. If Copilot is your daily driver, start there.
Grok Build & the open-source wave — the disruptors
Grok Build (launched July 8) is xAI’s entry: Grok 4.5, trained alongside Cursor’s team, priced aggressively at $2/$6 per million tokens, and available inside Cursor on all plans from day one. Ten days in, it’s too early for verdicts — but the price point alone will pull experimentation from every budget-conscious team.
Meanwhile the open-source side has quietly become the biggest story of the year. opencode passed 180k GitHub stars — more than any vendor CLI — by being MIT-licensed and completely model-agnostic: bring your own key, point it at Claude, GPT, Gemini, Grok, or a local GLM 5.2 (now the strongest open-weight coding model), and the harness is free. Cline, Aider, and Kilo CLI follow the same philosophy.
The portability layer — write once, run in every agent
The most practical development of 2026 isn’t any single tool — it’s that your customization now travels. Three conventions are supported nearly everywhere:
| Convention | What It Carries | Read By |
|---|---|---|
AGENTS.md | Repo-level rules: stack, conventions, build commands | Codex, Claude Code, Gemini, Copilot, opencode, Cursor |
SKILL.md | Reusable skills — instructions + scripts an agent can invoke | Claude Code, Codex, Gemini CLI, Copilot Agent Mode, Cursor |
| MCP servers | External tools: GitHub, Playwright, Postgres, Firecrawl… | Effectively all of them — same server config everywhere |
Which agent, for which task, today
Key takeaways
The harness is the product now
Frontier models converged; the wrapper decides your experience. Evaluate agents on context strategy, hooks, delegation, and token efficiency — not leaderboard decimals.
Run two, not one
The strongest pattern in real teams: one interactive agent for deep work (Claude Code or Codex CLI) plus one async agent for delegation (a cloud agent). They complement, not compete.
Make your config portable
AGENTS.md, SKILL.md, and MCP configs work across nearly every tool. Build your customization there once, and every agent you try next inherits it.
Watch the open-source line
When the most-starred agent on GitHub is MIT-licensed and model-agnostic, the pricing power of vendor CLIs has a ceiling. Grok Build’s $2/$6 pricing is the first shot; expect responses.