Claude Code vs Codex vs Gemini CLI vs Hermes: the October 2026 comparison

Codex vs Claude Code vs Gemini CLI? Claude Code suits an engineering team that wants the most built-in controls, Codex an organisation already on ChatGPT, and Gemini CLI a company equipped with Gemini Code Assist. All three read your repository, edit code and run commands.

What is the Google equivalent of Codex? Antigravity CLI for an individual, free with a Google account or included in Google AI Pro and Ultra. In a company, Gemini CLI remains available with a Code Assist licence, a paid API key or Vertex AI.

A command-line (CLI) coding agent reads your repository, edits files and runs commands in your terminal until the task is done.

Four vendors lead the field in October 2026: Anthropic with Claude Code, OpenAI with Codex, Google with Gemini CLI and its consumer successor Antigravity CLI, and NousResearch with Hermes Agent. OpenClaw completes the picture on the community side.

This comparison draws on months of use in real repositories. There is no absolute ranking: the best tool depends on your team, your budget and your data constraints.

Quick verdict

Claude Code is our default recommendation for an engineering team: Opus 5.5 on every plan, hooks, skills, per-tool approvals and a classifier-based auto mode. Codex is the natural pick for an organisation already on ChatGPT: it is included in every plan and its code is open. At Google, the answer depends on the account: Antigravity CLI for an individual, Gemini CLI for a company. Hermes Agent suits one person who wants a multi-model agent that learns from their sessions.

Agent	Entry price	Default model	Licence	Enterprise access
Claude Code	Pro $17/mo (annual) or $20	Opus 5.5	Proprietary	Team (Standard seat $20), Enterprise
Codex	Plus $20/mo for the CLI	GPT-6.1 Sol	Apache 2.0	Business ($20 or $25/user), Enterprise, Edu
Gemini CLI	Code Assist licence or paid API key	Gemini 3 family (3.1 Pro in preview), not optimized for Gemini 4 Argon	Apache 2.0	Code Assist Standard or Enterprise, Vertex AI
Antigravity CLI	Google account, Google AI Pro or Ultra	Not published (Gemini 4 Argon available)	Proprietary	Google Cloud projects
Hermes Agent	Free, API fees for the chosen model	Your choice (OpenRouter, Nous Portal, local)	MIT	Self-hosted
OpenClaw	Free, API fees for the chosen model	Your choice	MIT	Self-hosted, needs hardening

Prices come from the vendors’ official pages, in US dollars before tax, checked on October 16, 2026.

Rolling out a coding agent to a team? Colombani.ai trains developers on the tool you choose: our Claude Code training, two days on your repositories, or Codex for teams already on ChatGPT.

What changed in October 2026

The facts that changed this comparison since September, checked against the vendors’ changelogs and pricing pages.

• Gemini CLI is no longer Google’s free agent. Since June 18, 2026, free accounts and Google AI Pro and Ultra subscribers move to Antigravity CLI. Gemini CLI remains available to Code Assist Standard or Enterprise licence holders, paid API keys and Vertex AI.
• Claude Code switched to Opus 5.5 by default in version 2.1.280 (September 22), including on Pro and Team Standard. Sonnet 5.5 and Haiku 5.5 followed, all with a one million token context window.
• Codex moved to a new model generation: GPT-6 Luna and GPT-6 Sol on September 22, GPT-6.1 Sol on September 29. GPT-5.5 was retired on October 14 for users signed in with their ChatGPT account.
• Codex is built into the ChatGPT app since July 9, 2026 (version 26.707), and the codex mcp-server command was removed on September 5.
• Hermes Agent ships stable releases: 0.22.0 (October 14) installs tested releases by default, runs local models in one click and opens a 600-plugin catalog.
• Versions on October 16: Claude Code 2.1.301 and Codex 0.163.0 (Oct 15), Gemini CLI 0.64.0 (Oct 13), Antigravity CLI 1.3.6 (Oct 15), Hermes Agent 0.22.2 (Oct 15).

How this comparison was evaluated

The verdict rests on repeated use in real repositories rather than a one-shot benchmark. Five criteria matter in production: quality on multi-file changes, context retention, tool integration, team governance and total operating cost.

Every price, version and licence quoted here was checked on October 16, 2026 against the vendor’s documentation, GitHub repository or changelog, linked at the end. Colombani.ai trains teams on Claude Code and on Codex, so readers know where the author stands.

The table

All of these agents support MCP (Model Context Protocol), the standard protocol that lets them plug into the same external tools: databases, ticketing, documentation. The real differences lie elsewhere.

	Claude Code	Codex	Gemini CLI	Antigravity CLI	Hermes Agent	OpenClaw
Maker	Anthropic	OpenAI	Google	Google	NousResearch	Community
Default model	Opus 5.5	GPT-6.1 Sol (GPT-6 Luna on Free and Go)	Gemini 3 family	Not published; Gemini 4 Argon, Claude Opus 5.5 and Sonnet 5.5 through a Google Cloud project	Your choice	Your choice
Context window	1M tokens	Depends on model	1M tokens	Depends on model	Depends on model	Depends on model
Included in which plans	Pro, Max, Team, Enterprise	Every ChatGPT plan	Code Assist Standard and Enterprise	Google account, Google AI Pro and Ultra, Google Cloud	None, API fees	None, API fees
CLI available from	Pro, $17	Plus, $20	Code Assist licence or paid API key	Google account	Free	Free
Free install and free plan	No	Yes, desktop app only	No for individuals	Yes, quotas not published	Yes, free software, API fees	Yes, free software, API fees
Licence	Proprietary	Apache 2.0	Apache 2.0	Proprietary	MIT	MIT
Version (Oct 16)	2.1.301	0.163.0	0.64.0	1.3.6	0.22.2	2026.10.4
MCP	Yes	Yes	Yes	Yes	Yes	Partial
Multi-agent	Stable subagents, experimental agent teams	Native subagents	Subagents	Subagents, background tasks	Agent-to-agent delegation (A2A)	No
Customization	Skills, Hooks, plugins	Skills, hooks, plugins	Skills, Hooks, extensions	Skills, Hooks, plugins	Auto-generated skills, 600 plugins	Community skills
Project memory	CLAUDE.md, or AGENTS.md as a fallback	AGENTS.md	GEMINI.md	Rules and persistent history	Automatic learning loop	Persistent local memory
Surfaces	Terminal, desktop, web, IDEs	Terminal, ChatGPT, IDEs, web, iOS	Terminal, IDEs	Terminal, Antigravity 2.0	Terminal, desktop, messaging apps	Terminal, messaging apps
Data residency	Bedrock, Vertex AI or Foundry; optional US-only inference	Residency controls on Enterprise; Amazon Bedrock	Vertex AI in your Google Cloud project	Not specified in public documentation	Your provider’s, local possible	Your provider’s

Data residency means where your requests are processed and stored. For an organisation subject to GDPR, this row often weighs more than the price.

Claude Code: the oldest and best-equipped

Claude Code is the oldest CLI agent from a major vendor: its public repository dates from February 2025, two months before those of Codex and Gemini CLI.

Since version 2.1.280 (September 22, 2026), Opus 5.5 is the default model on every plan, including Pro and the Team Standard seat. Opus excels at long tasks: multi-file refactoring, subtle bug hunting, system architecture.

The agentic loop. Claude Code sends a request to the Claude API. The response includes a stop_reason: "tool_use" means the model wants to use a tool (read a file, run a command), "end_turn" means it’s done. The model itself decides when to stop.

Agentic search. Anthropic abandoned the RAG approach (vector database) in favor of iterative search: Grep, Glob, Read. The model searches, refines, searches again. Slower than a vector index, but significantly more reliable on real code.

Auto mode uses classifiers that evaluate the risk of each action before executing it. Writing a test file: low risk, automatic execution. Deleting a directory: high risk, asks for confirmation.

The harness, meaning everything around the model that guides and constrains it. CLAUDE.md gives the model persistent context: coding conventions, architecture, business constraints. Since version 2.1.277 (September 18), Claude Code reads AGENTS.md when a repository has no CLAUDE.md. Skills package reusable procedures, and Hooks run a check before or after each action.

Beyond the terminal. Subagents are stable and run in the background. Agent teams, where several instances split a task, remain experimental and off by default. Routines run scheduled tasks in the cloud, and since September 17 Projects coordinate parallel sessions with shared memory.

Opus is not the top of the range. Anthropic also offers Fable, available in Claude Code through the best alias, with its own weekly limit and, depending on the plan, usage credits.

The downsides. Claude Code is not open source and has no free plan. The one million token context window is now standard on Opus 5.5, Sonnet 5.5 and Haiku 5.5, so the old context limitation is gone.

Codex: open source and built into ChatGPT

Codex is OpenAI’s agent, written in Rust. Its code has been published under the Apache 2.0 licence since the first commit on April 16, 2025. This permissive licence allows commercial use, modification and redistribution, and expressly grants a patent licence.

Built into ChatGPT. Since July 9, 2026 (version 26.707), Codex is part of the ChatGPT desktop app on macOS and Windows. For an organisation already on ChatGPT Business or Enterprise, the coding agent is available without a new contract.

The models. GPT-6 Luna and GPT-6 Sol arrived on September 22, 2026, and GPT-6.1 Sol on September 29. GPT-5.5 was retired on October 14 for users signed in with their ChatGPT account. Sessions authenticated with an API key are not affected.

What each plan gives you. Free and Go ($8) open Codex in the desktop app with GPT-6 Luna. The CLI, the IDE extension and Codex on the web are available from the Plus plan ($20), with GPT-6.1 Sol. Pro comes at $100, $200 and $500, Business costs $20 per user billed annually or $25 monthly, and Enterprise and Edu are priced on request.

Native subagents and fast startup. Codex delegates subtasks to parallel agents. Charlie Marsh, the author of uv and ruff, rewrote its startup phase on August 20, which is now about 25 times faster. Amazon Bedrock is natively supported as a provider, useful when code must stay on infrastructure you control.

The downsides. The codex mcp-server command, which exposed Codex as an MCP server to other tools, was removed on September 5, 2026: integrations that used it must move to the app server. Usage limits depend on the plan and the model, and are best measured on the usage dashboard rather than the price list.

Claude Code vs Codex: which one for a team

This question comes up constantly, and the two tools have converged: repository access, MCP, skills, subagents, hooks. Comparing their features is no longer enough to tell them apart. Three questions settle it.

Which subscription does the organisation already pay for? A company on ChatGPT Business gets Codex at no extra cost. A company on Claude Team gets Claude Code from the Standard seat. The entry price is the same: $20 a month billed annually.

What does it need to be able to verify? Codex publishes its harness code under Apache 2.0, which makes it auditable. Claude Code keeps a readable local record of what the agent saw and sent, with tool-by-tool approvals.

Who reviews the code it produces? For a team shipping to production, Claude Code with Opus 5.5 remains our default recommendation. Codex takes the lead when internal policy requires an open licence or routing through Amazon Bedrock.

The two combine well: one agent writes, the other reviews, and their blind spots differ. The detailed analysis is in Codex or Claude Code: how to choose in 2026.

Gemini CLI and Antigravity CLI: the Gemini equivalent of Codex

This is the most frequent question behind this page, and the answer changed in June. On May 19, 2026, Google announced the transition from Gemini CLI to Antigravity CLI. Since June 18, Gemini CLI is no longer available to free accounts or to Google AI Pro and Ultra subscribers.

For an individual: Antigravity CLI. Available to everyone, it shares its engine with Antigravity 2.0, Google’s desktop app for agent-first development. It is written in Go and orchestrates several agents in the background. It keeps Gemini CLI’s key features: Skills, Hooks, subagents and extensions, now called plugins.

For a company: Gemini CLI remains available. Organisations with a Gemini Code Assist Standard or Enterprise licence keep their access, as do paid API keys and Vertex AI. Version 0.64.0 (October 13, 2026) runs on the Gemini 3 family, with Gemini 3.1 Pro in preview. The repository README states that Gemini CLI has not been optimized for Gemini 4 Argon and points to Antigravity for that model.

What sets them apart. Gemini CLI stays open source under Apache 2.0. The Antigravity CLI repository only publishes documentation and the changelog. Antigravity CLI, on the other hand, offers Claude Opus 5.5 and Sonnet 5.5 to authorised Google Cloud projects. Its auto mode sends any command that would leave the sandbox to a second agent for review.

What is no longer true. Gemini CLI is no longer “free with a Google account”. The repository README now says so at the top: for the unpaid tier and Google One users, Gemini CLI was replaced by Antigravity CLI on June 18, 2026. Gemini CLI also has Skills, Hooks and subagents, contrary to what this comparison stated before October. Gemini Code Assist for GitHub has accepted no new installations since June 18, except through Google Cloud.

Antigravity CLI’s free-tier quotas do not appear in the sources we checked, so we do not quote them. For Gemini on the productivity side, see Gemini for business.

Hermes Agent: the one that learns from you

Hermes Agent from NousResearch is the most original on this list. Its standout feature: a built-in learning loop. The agent creates skills from your sessions, improves them through use, and builds a persistent profile of your preferences.

NousResearch describes it as the only agent with a built-in learning loop: it adapts to its user automatically, where Claude Code requires maintaining a CLAUDE.md file by hand.

Multi-provider. Hermes works with OpenRouter, Nous Portal, OpenAI, your own endpoint or local models, and can switch models mid-session. Since 0.22.0, it installs llama.cpp and runs an open model on your machine in one click, with no account and no API key.

Multi-platform. Beyond the terminal, Hermes can be driven from a desktop app, Telegram, Discord, Slack, WhatsApp or Signal, through a single gateway.

A project that has grown considerably. Six months after its first public releases in March 2026, Hermes has around 254,000 GitHub stars as of October 16, under the MIT licence. The 0.22.0 cycle alone brought together about 2,500 contributors. Version 0.22.2 (October 15) is the latest stable release.

The downsides. The pace is still dizzying: about 24,000 commits between 0.21.0 and 0.22.0. The stable release channels introduced on October 14 reduce the risk, but pinning a version remains the rule before any client work. The learning loop can retain mistakes. And the optional GODMODE skill, which ships sample prompts for bypassing guardrails, raises a legitimate security question.

Hermes vs Codex: two philosophies of agent

Codex is a software vendor’s coding agent, tied to a ChatGPT subscription and to OpenAI models. Hermes is a provider-independent personal agent that also writes code.

Choose Codex when the agent serves a team that must share configuration, billing and administration: ChatGPT workspace, SSO, audit logs, residency controls on Enterprise.

Choose Hermes when one person wants a single agent for code, messaging and scheduled tasks, with the model of their choice, including a local one. The hermes import-agent command, added in August, imports an existing Codex or Claude Code setup.

Both licences are permissive: Apache 2.0 for Codex, MIT for Hermes. The difference is governance: Codex is administered like an enterprise service, while Hermes is self-hosted and secured by its user.

OpenClaw: the viral phenomenon

OpenClaw (ex-Clawdbot, ex-Moltbot) has the most chaotic trajectory of the bunch: renamed twice after an Anthropic trademark complaint, 60,000 GitHub stars in 72 hours at launch, around 392,000 as of October 16, 2026, and its creator, Peter Steinberger, hired by OpenAI.

Persistent local memory. This feature is what made it a success. OpenClaw remembers your sessions for weeks, with no configuration.

Works with several models. It runs with Claude, GPT or open models such as Kimi.

More than a coding agent. OpenClaw positions itself as a “Life OS”: WhatsApp connection, calendar management, third-party app control. The most ambitious approach, and the riskiest.

The downsides. Security. 923 OpenClaw gateways were recorded as exposed without authentication, with shell access and API keys in plain text. If you deploy OpenClaw, lock it down before any use.

Other terminal agents

GitHub Copilot CLI requires an active Copilot subscription and can be disabled by the organisation’s administrator (version 1.0.99, October 16). OpenCode is an open source, multi-model agent under the MIT licence, with around 214,000 stars (version 1.19.1, October 15). Amp runs its agents remotely: they keep working after you close your laptop. Cursor CLI brings Cursor’s agent to the terminal, with models from Anthropic, OpenAI, Google and Cursor.

Already deploying agents in production? Beyond the tool, two turnkey offerings:

• Business AI Agent Plugins: skills, hooks and MCP assembled, tested and evaluated, source code delivered. Built with an agentic systems architect method backed by the Claude Certified Architect Professional certification, issued by Anthropic.
• AI Security Audits: pentest, threat modeling and vulnerability detection accelerated by AI agents. OWASP, MITRE ATT&CK, CIS frameworks. Covered by Colombani.ai professional liability insurance.

Which agent for which profile?

Here are our recommendations by use case.

Demanding solo developer: Claude Code. Opus 5.5 by default from the Pro plan makes the difference when every mistake costs you time. Alternative: Antigravity CLI to get started with a Google account.

Structured engineering team: Claude Code. Shared CLAUDE.md in Git, Hooks to enforce conventions, worktrees for parallel work. The investment in the harness pays off once the team exceeds two or three people.

Organisation already on ChatGPT Business or Enterprise: Codex. The agent is included, so is the administration, and the Apache 2.0 licence makes the harness auditable.

Company on Google Cloud: Gemini CLI with a Code Assist licence, or Antigravity CLI attached to your Google Cloud projects.

Enterprise or mid-size company: Claude Code. Governance (Skills, security Hooks, permissions) and the tool’s track record meet the requirements of structured teams.

Team building its own agents: once the coding agent is mastered, the MCP and Agent SDK training teaches how to design, evaluate and run custom agents, with their MCP servers.

Student or learner: Antigravity CLI with a Google account, or Codex in the ChatGPT app from the free plan.

Open source contributor: Codex, Gemini CLI, Hermes Agent or OpenCode, all published under open licences.

Multi-model experimentation: Hermes Agent. Mid-session model switching and one-click local models make it easy to compare LLMs on the same project.

Power user: OpenClaw. Persistent memory and the multi-platform approach are unique. Secure your installation.

Business teams who do not code: these command-line agents are not the right entry point; the AI agents training for companies runs on the tool already in place, with no terminal; for an organisation on ChatGPT, the ChatGPT training for companies follows the same approach.

To roll out one of these agents to a team, Colombani.ai offers two courses: Claude Code in production and Codex.

Our pick (and why)

At Colombani.ai, we train first on Claude Code, because it brings together the most complete set of tools: CLAUDE.md, Skills, Hooks, per-tool approvals and a classifier-based auto mode. We also train on Codex for teams that work in ChatGPT.

Every training shows the alternatives, because methodology matters as much as the tool. Structuring your context, planning before coding, verifying systematically: these principles work with any agent.

A well-written CLAUDE.md is a project conventions document, and Claude Code now reads AGENTS.md, the format Codex also uses. Skills exist at Anthropic, OpenAI and Google. The discipline of verification (tests, diff review) is universal.

The field moves fast: Google replaced its consumer agent within a month, OpenAI changed models twice in one week, and Hermes passed 250,000 stars in six months. A structured method for coding with AI keeps all its value.

Once the tool is chosen, the Claude Code in production or Codex course gets your team to adopt the method, on your own repositories.

Sources

• Claude Code: changelog, Anthropic
• Claude plans: Anthropic
• Codex: GitHub repository, OpenAI
• Codex: pricing and plans, OpenAI
• Codex: changelog, OpenAI
• Transitioning Gemini CLI to Antigravity CLI: Google Developers Blog, May 19, 2026
• Gemini CLI: GitHub repository, Google
• Antigravity CLI: GitHub repository, Google
• Hermes Agent: NousResearch
• Harness design for long-running apps: Anthropic Engineering