Frontier models on a subscription: Claude, ChatGPT, Grok, Gemini
Four vendors, four monthly plans, four terminal coding agents. What each tier unlocks, what it costs, how to hook it up on the mini-PC, and what each vendor really allows.
In this guide
Guide checked 3 months ago: some commands may have changed. Let us know if so.
In short
The four big vendors each sell a subscription that unlocks their terminal coding agent: Claude Pro ($20 a month) or Max ($100 or $200) for Claude Code, ChatGPT Plus ($20) or Pro ($100 to $500) for Codex, SuperGrok ($30) for Grok Build, Google AI Pro (€21.99) or Ultra (€99.99 or €219.99) for Antigravity CLI, prices checked on October 1, 2026. The right way to work is hybrid: the mini-PC runs the agents, one or two frontier plans do the bulk of the reasoning, and Ollama's local models take the confidential, the repetitive and the offline. The rules differ: OpenAI and xAI let you plug their subscription into OpenCode, Hermes or OpenClaw, Anthropic keeps it for Claude Code (which those tools can drive), and Google forbids it outside its own tools.
Do this first: Why a mini-machine
Your mini-PC is a great workshop. Still, it will never run a model of the Claude Opus or GPT-6 class at home: those frontier models live in their vendors’ data centers. The good news is that you can buy them by the month, like a music subscription, and every vendor ships a coding agent with its plan that runs right in your terminal.
This guide goes through all four: Anthropic (Claude Code), OpenAI (Codex), xAI (Grok Build) and Google (Antigravity CLI). Prices, limits and usage rules were checked on the official pages on October 1, 2026. They change often: the date matters.
Why hybrid is the right way to work
Three players, three roles:
- The mini-PC orchestrates. It runs the agents and holds your projects, your Git repositories, your
tmuxsessions and your services. It’s always on, reachable from anywhere through Tailscale, and it only costs its electricity. - Frontier models do the reasoning. Architecture, big refactors, nasty debugging, review: that’s where the gap with local models really shows. On a subscription, they cost a fixed amount every month instead of a per-token bill that explodes on your first weekend of intense coding.
- Local models take the rest. Code or data that must not leave the house, repetitive bulk tasks (sorting, summarizing, rewording a thousand files), and offline work. Free to use, private by design (see Ollama & local models).
The Going hybrid: cloud + local guide shows how to make both worlds work together. This one helps you pick the cloud plan or plans.
The subscriptions, side by side
Checked on October 1, 2026, on each vendor’s page. Dollar prices are the ones shown on the Anthropic (before tax, even on the French page), OpenAI and xAI pages; Google shows its French prices in euros.
| Vendor | Tier | Price per month | Coding agent | What the page says about limits |
|---|---|---|---|---|
| Anthropic | Pro | $20 ($17 billed annually) | Claude Code | At least 5 times the free usage per 5-hour session, plus weekly limits. The quota is shared between claude.ai and Claude Code. |
| Max 5x / Max 20x | $100 / $200 | Claude Code | 5 or 20 times Pro usage per 5-hour session, with a weekly limit across all models. Priority access at peak times. | |
| OpenAI | Go | $8 | Codex, in the desktop app (rolling out) | GPT-6 Luna only. |
| Plus | $20 | Codex on the web, in the CLI, in the IDE extension and on iOS | OpenAI’s estimate per 5-hour window: 5 to 45 messages with GPT-6 Astra, 15 to 160 with GPT-6.1 Sol. Weekly limits may also apply. | |
| Pro | $100, $200 or $500 | Codex | No 5-hour limit “currently”. GPT-6 Astra’s Ultrafast mode is reserved for Pro $500. | |
| xAI | SuperGrok | $30 | Grok Build | Higher limits than the free plan, no published figure. |
| SuperGrok Plus | $100 | Grok Build | ”Significantly higher usage” across Chat, Imagine, Voice and Build. | |
| AI Pro | €21.99 | Antigravity and Antigravity CLI | Quota refreshed every 5 hours, up to a weekly limit. AI credits to buy for overages. | |
| AI Ultra | €99.99 or €219.99 | Antigravity and Antigravity CLI | Two levels: 5 or 20 times AI Pro’s limits, and in Antigravity “higher” or “the highest” limits. |
Sources: Claude pricing, Max plan, Codex pricing, xAI pricing, Google AI subscriptions, Antigravity plans.
A few details that matter:
- Free. Claude Code isn’t included in Claude’s free plan. Codex is (GPT-6 Luna, in the desktop app, rolling out). Grok Build is ticked for every xAI plan, the free one included, with limits that rise with the subscription. Antigravity has a free tier with a weekly quota.
- Included models. Claude Pro opens Opus, Sonnet and Haiku; Fable, the most expensive, only comes through paid usage credits, whereas with Max it can use up to half of the weekly limits. Codex offers GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna (GPT-5.5 leaves Codex on October 14, 2026). Antigravity serves Gemini 3.8 Flash, 3.7 Flash, 3.6 Flash, 3.1 Pro, Claude Sonnet and Opus 4.6 and gpt-oss-120b; its plans page, though, reserves third-party models for AI Ultra, a contradiction Google hasn’t cleared up.
- Tiers missing from the table. xAI also sells SuperGrok Lite and SuperGrok Heavy, whose price the public page doesn’t show. Google AI Plus (€4.99) isn’t listed for Antigravity.
Tool
What does the setup cost per month?
Tick your subscriptions, pick the machine, and you get a month's bill: the electricity of a machine on 24/7, plus the plans.
Per month
- Electricity18.9 kWh, 24/7 with 2 h under load per day
- €4.74
- Subscriptions
- $20
- Total
- €4.74 + $20
Enter the idle and load power draw to work out the electricity.
Dollars stay in dollars: no conversion, since the rate and taxes depend on your bank and country.
No estimate for per-token APIs: we have no reliable data on what you will use. To compare model prices with their quality, see the Quelle IA value-for-money ranking (in French).
What the models behind them are worth
In Quelle IA’s September 28, 2026 edition, Claude Opus 5.5 leads the Code ranking (99.9), statistically tied with four other models, including Claude Fable 5.1, Claude Sonnet 5.5 and GPT-6 Astra. GPT-6 Astra leads the Agents ranking, tied with three others, including Claude Fable 5.1 and Claude Opus 5. Grok 4.6 and 4.7 follow at 91.0 in code, Gemini 3.8 Flash at 90.5. In other words, Anthropic and OpenAI share the top, and xAI and Google trail by about ten points. To put two models head to head, there’s Compare; for score versus price, the value-for-money ranking (all in French).
Vendor by vendor: installing the agent on the mini-PC
All four agents install through a script, with no Node.js. The mini-PC often has no screen or browser: each one offers a sign-in where you open a URL on your computer and copy a code.
Anthropic: Claude Code
The most battle-tested agent on long tasks, and the one this site has followed from day one.
# Installs Claude Code into ~/.local/bin
curl -fsSL https://claude.ai/install.sh | bash
# First launch: sign in to your Claude account (Pro or Max)
claude
Over SSH, the browser doesn’t open: copy the login URL (c key), open it on your computer, then paste into the terminal the code the browser shows. Details in Installing the agent.
OpenAI: Codex
Codex comes as a CLI, an editor extension, on the web and in the ChatGPT app, and they all share the same quota.
# Installs the Codex CLI (official installer)
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# On a machine with no browser: device code sign-in (beta)
codex login --device-auth
# Then, in a project folder
codex
Device code sign-in must first be enabled in your ChatGPT account’s security settings. Otherwise, sign in on your computer with codex login, then copy the ~/.codex/auth.json file to the mini-PC: it holds your access tokens, so treat it like a password. Inside a session, /status shows what’s left of the quota.
xAI: Grok Build
xAI’s coding agent, launched in beta in May 2026 and opened to all SuperGrok subscribers in early June. It reads your AGENTS.md, your skills and your MCP servers, and can work in plan mode.
# Installs Grok Build (the command is called grok)
curl -fsSL https://x.ai/cli/install.sh | bash
# Device code sign-in, for a machine with no browser
grok login --device-auth
# Then, in a project folder
cd my-project && grok
Update: grok update. Inside a session, /model switches models. The in-house model, Grok Build 0.1, only has a provisional score on Quelle IA so far (79.8 in code).
Google: Antigravity CLI
Since June 18, 2026, Gemini CLI no longer serves free accounts or Google AI Pro and Ultra: its successor for these subscriptions is Antigravity CLI, whose command is agy.
# Installs Antigravity CLI
curl -fsSL https://antigravity.google/cli/install.sh | bash
# First launch: sign in to your Google account
agy
Over SSH, agy detects it and shows an authorization URL to open in your computer’s browser, then asks you for the code you get. A paid Gemini API key works too, but it’s billed per use, outside the subscription.
Which subscription works where?
A subscription works in its vendor’s own agent. Some vendors go further and let other tools use it. Here’s what you can do with the five tools this site follows: the three coding agents (Claude Code, Codex, OpenCode) and the two always-on agents (Hermes and OpenClaw).
| Subscription | Claude Code | Codex | OpenCode | Hermes | OpenClaw |
|---|---|---|---|---|---|
| Claude Pro / Max | ✅ in-house tool | ❌ | ❌ forbidden by Anthropic | ⚠️ sign-in offered, which only uses usage credits bought on top of a Max plan; better to have it drive claude | ✅ by driving the installed claude program |
| ChatGPT Plus / Pro | ❌ | ✅ in-house tool | ✅ | ✅ | ✅ |
| SuperGrok / X Premium | ❌ | ❌ | ✅ | ✅ | ✅ |
| Google AI Pro / Ultra | ❌ | ❌ | ❌ forbidden by Google | ❌ | ❌ forbidden by Google |
| OpenCode Go | ✅ | ✅ | ✅ in-house tool | ✅ | not listed |
| Ollama Cloud | ✅ | ✅ | ✅ | not listed | not listed |
Sources, checked on October 1, 2026: Claude Code terms, OpenCode providers, Sign in with ChatGPT, xAI announcements for OpenCode, Hermes and OpenClaw, Antigravity terms, Hermes providers, OpenClaw authentication. The OpenCode Go and Ollama Cloud rows list the tools named on each offer’s page, as recorded by Quelle IA (in French) on September 30; “not listed” means the page doesn’t mention it.
Hermes and OpenClaw can also drive the agents themselves: they launch claude or codex on the mini-PC, each signed in to its own subscription. That’s the safest route for a Claude subscription, since Anthropic’s program runs unmodified.
Combining several
Nothing forces you to pick just one. A common setup: a Claude or ChatGPT plan for the bulk of the work, a local model for the rest, and possibly a second plan to compare or to take over when the first hits its limit.
OpenCode, the power strip
OpenCode accepts subscriptions, API keys and your local models all at once, and you switch models mid-session. An example with ChatGPT, SuperGrok and Ollama:
0 of 4 steps done Your ticks stay in this browser.
-
Install OpenCode
curl -fsSL https://opencode.ai/install | bash opencode -
Plug in the subscriptions
In OpenCode, type
/connect, choose OpenAI then the ChatGPT Plus/Pro option. Do it again with xAI and the SuperGrok Subscription option, which works with a device code and so suits the mini-PC. GitHub Copilot and Z.ai’s GLM Coding Plan plug in the same way. -
Add the local model
If Ollama is already running,
ollama launch opencodestarts OpenCode on a local model, without touching your configuration. -
Pick the engine for the task
/modelslists everything that’s plugged in. Hard reasoning goes to GPT or Grok, confidential and repetitive work to the local model.
Claude, for its part, only gets into OpenCode through an Anthropic API key, billed per use. If your main plan is Claude, keep Claude Code in one terminal and OpenCode in another: tmux lets them live side by side with no fuss.
Hermes and OpenClaw, for always-on agents
For an agent that runs all the time and answers on Telegram, both gateways connect to the subscriptions that allow it:
# Hermes: choose the ChatGPT (Codex) subscription or "xAI Grok OAuth"
hermes model
# OpenClaw: ChatGPT subscription, then SuperGrok, with device code sign-in
openclaw models auth login --provider openai --device-code
openclaw models auth login --provider xai --method oauth
# OpenClaw: route Claude models through the already signed-in claude program
openclaw models auth login --provider anthropic --method cli --set-default
Everything else (install, security, messaging apps) is in Hermes & OpenClaw.
Multi-tool plans and routers
Other plans are sold to be plugged in anywhere: OpenCode Go ($10 or $40), Ollama Cloud ($20 or $100), Z.ai’s GLM Coding Plan ($18 to $168), Kimi Code ($19 to $199). They give access to open models, cheaper and a notch below the frontier in code. Quelle IA keeps the list up to date, with what each tier counts: coding subscriptions (in French). And if you’d rather pay per use, a gateway like OpenRouter or OpenCode Zen gives you a single key for hundreds of models (see Installing the agent).
Where to start
- One plan, for coding every day: Claude Pro with Claude Code, or ChatGPT Plus with Codex. Both lead the rankings, at $20 a month.
- You want a single tool for everything: ChatGPT Plus or SuperGrok in OpenCode, plus your local models.
- You hit the limits every week: move up a tier (Max or Pro), or add a second vendor’s plan.
- You already have Google AI Pro or an X Premium subscription: try Antigravity CLI or Grok Build before paying for anything else.
All the commands in this guide
Frequently asked questions
Subscription or API key: which should you choose for coding with AI?
For coding every day, a subscription is usually the cheapest option: a fixed price and usage caps, often counted per 5-hour window and per week, designed for you at the keyboard. An API key is paid per token consumed, with no cap other than the one you set, and it is the intended route for shared automation, products and servers.
Can you use a coding agent for free?
Partly, according to the pages checked on 1 October 2026. Claude Code is not included in Claude's free plan, whereas Codex is, with GPT-6 Luna in the desktop app, rolling out gradually. Grok Build is open to every xAI plan, the free one included, and Antigravity has a free tier with a weekly quota.
Can you run scripts with a Claude subscription?
Yes, to a point: since the 15 June 2026 update, claude -p and the Agent SDK draw on your plan, and claude setup-token creates a one-year token for scripts. The limits do, however, assume ordinary, individual use. For a shared service or a product, Anthropic points you to an API key.
How do you sign a coding agent in on a mini-PC with no browser?
Each agent offers a sign-in where you open an address on your computer and copy back a code. In Claude Code launched over SSH, the c key copies the sign-in address, then you paste into the terminal the code the browser displays. Codex and Grok Build have a login --device-auth command, and Antigravity CLI detects the SSH session on its own.
Which subscription should you pick for coding every day?
One plan is enough to start: Claude Pro with Claude Code, or ChatGPT Plus with Codex, at $20 a month each, since Anthropic and OpenAI share the top of the Quelle IA rankings. If you hit the limits every week, move up a tier or add a second provider's plan. If you already have Google AI Pro or X Premium, try Antigravity CLI or Grok Build first.
Terms in this guide: Frontier modelAgentClaude CodeCodexGrok BuildAntigravity CLIGitTokenCLISSHSkillMCPAPI keyOpenCodeHermesOpenClawUnified APIOllamaCoding plan
Spotted a mistake?
A command stopped working, a price changed?
Tools change every month. Tell me what is wrong in this chapter and I will fix it and update its date.
