Gemini 4 Argon for Coding Agents: Setup, Costs, and Which Tasks to Route to It

Gemini 4 Argon has been generally available since October 13, 2026, as gemini-4-argon, a stable model on the paid tier of the Gemini API. Google AI Pro and Ultra subscribers also get it in the Gemini app, and Ultra subscribers in Antigravity. There is no free API tier. For a Gemini 4 Argon coding agent, the confirmed numbers are a 1,048,576-token context window shared by input, thinking, and output; 65,536 output tokens per request by default, up to about 1M in a background mode; a lead on the DeepSWE v1.1 coding benchmark at 77.9%; losses on Terminal-bench 4.0 and FrontierSWE v2; and $2 per million input tokens and $10 per million output through November 30, then $4 and $20 from December 1.

The tension is no longer access. It is what a new frontier model does to your board: it tempts a team to rebuild around it, and the result is a week where nobody is sure which agent is on what. If a Gemini Agent is already running real Tasks on a model you trust, Argon is a one-value change on one Computer, followed by one Task you can compare against the last. One caveat shapes this guide: Google has not said Gemini CLI supports Argon, so you confirm it on your own Computer before Argon gets a Task.

TL;DR

Install Gemini CLI on a connected Computer and authenticate it with a paid Gemini API key, a Gemini Enterprise Agent Platform key, or a Code Assist Standard or Enterprise license; consumer Google accounts lost Gemini CLI on June 18, and a Pro or Ultra subscription is not an API key. Confirm the Runtime reads Available, create an Agent on it, and pin the model to gemini-3.8-flash. Then change one value to gemini-4-argon, run a one-line smoke prompt, rescan, run one Task you already ran, and compare. Route by benchmark row, not by headline.

What’s confirmed, and what isn’t

As of October 15, 2026: Google announced Argon on September 30 for a set of cyber-defense partners in its Fairwind Program, and the Gemini API changelog lists gemini-4-argon as generally available on October 13. Google’s pricing page says Argon is not available on the Free tier; Batch and Flex inference cost half the standard rate.

The rows that matter for a coding crew, from Google’s own evaluation table:

Read it with the caveat Google’s methodology prints: Google computed Argon’s DeepSWE and Terminal-bench scores itself, while most competitor numbers are self-reported by their labs or taken from public leaderboards. Harnesses differ. Since launch, Cursor has reported 52.8% for Argon on its own CursorBench at default effort, which it calls on par with Opus 5.5. Our three-way comparison goes row by row.

Three spec details change how you run it. The 1M window is shared, so a run that reads a large repository has less room left to write. A synchronous request stops at 65,536 output tokens by default; the full output needs the Interactions API’s background long decode, which can take minutes. And Argon drops explicit caching and ignores temperature, so old Gemini 3 settings carrying either need cleaning. Still unpublished: the knowledge cutoff, Vertex AI availability, and any statement about Gemini CLI.

Step 1: a Gemini CLI Runtime on a Computer

This step has a trap that predates Argon. On June 18, Gemini CLI stopped serving Google AI Pro, Google AI Ultra, and free Code Assist for individuals accounts, and Google moved those users to Antigravity CLI. Per Google’s transition notice, Gemini CLI remains available through paid Gemini API keys, Gemini Enterprise Agent Platform keys, and Code Assist Standard or Enterprise licenses. A personal Google login no longer gives you a working Runtime, and neither does a Pro or Ultra subscription, even though both now include Argon in the Gemini app. A paid key is the simplest route for unattended runs, and Argon needs one anyway.

On the Computer that will run the work, install or update Gemini CLI and confirm it works in a plain terminal first:

export GEMINI_API_KEY="your-paid-project-key"
gemini -m gemini-3.8-flash -p "Reply with the word ok."

If that returns, the tool works on that host. Open Computer detail and run the rescan or connection test; the Gemini CLI Runtime should read Available. If it shows Detection failed or reports that it is not signed in, follow the detail page’s guidance, then rescan. Computers in Sharkly covers the statuses.

Step 2: an Agent, and where the model actually lives

Sharkly does not pick models. The Agents documentation is direct: “The Agent follows the Runtime’s default model. It does not promise or display a specific model name.” To change models, the docs point to Custom arguments on the Agent or Runtime, or to the Runtime or Computer tool configuration, and there is no model dropdown on the Agent.

The Agent defines how work should be handled; the Computer supplies the host; the Runtime, Gemini CLI here, performs the agent session; the Task remains the shared record. Create the Agent from Agents, New Agent: pick the Computer and the Gemini CLI Runtime, set the default working directory to Create temporary directory so each Task gets an isolated directory and, for repository-backed runs, a fresh worktree, and keep the instructions short. Agents in Sharkly covers the other run settings.

The model is a Gemini CLI setting. Gemini CLI resolves it in a documented order: the -m or --model flag, then the GEMINI_MODEL environment variable, then model.name in settings.json, then the default, auto. Gemini CLI’s user settings file on the Computer, ~/.gemini/settings.json, covers every Gemini CLI session for that user:

{ "model": { "name": "gemini-3.8-flash" } }

# Agent environment variable: only this Agent's runs
GEMINI_MODEL=gemini-3.8-flash

# or an Agent Custom argument (additional Runtime arguments)
-m gemini-3.8-flash

The settings file is the safe default. The Agent-level routes help when one Computer serves several Agents and only one should move; test them on one Agent before standardizing. Pin the model rather than leaving auto, which routes between models and can fall back to another on quota errors, ruining any comparison.

Step 3: switch one Agent to Argon

Start with the API, so a key problem never looks like a CLI problem. This is Google’s own request, sent from the Computer:

curl "https://generativelanguage.googleapis.com/v1beta/interactions" \
  -H "x-goog-api-key: $GEMINI_API_KEY" \
  -H 'Content-Type: application/json' \
  -X POST \
  -d '{
    "model": "gemini-4-argon",
    "input": "Your prompt here",
    "generation_config": {"thinking_level": "medium"}
  }'

If that answers, the key can call Argon (thinking_level takes low, medium, which is the default, or high). Then try Gemini CLI:

gemini -m gemini-4-argon -p "Reply with the word ok."

That smoke prompt is your evidence, not this article. Google has not said Gemini CLI supports Argon, and as of October 1 its source had no Argon entry. If the prompt fails, update Gemini CLI (sharkly computer update requests a CLI update for a specific Runtime) and try again. If it still fails, the Agent stays on Flash and nothing on the board breaks.

If it passes, change one value: model.name, GEMINI_MODEL, or the -m argument. Rescan from Computer detail and confirm Available. Then assign one Task you already ran on Flash and compare the two results where they return, on the Task: the diff, the checks the Agent ran, and what the reviewer had to fix.

When Argon gets the Task, and when it doesn’t

Give Argon the Task when the work means reading a lot to change a little, or producing one long change that has to hold together: a cross-module issue in a large repository, a migration across thousands of lines. Google’s table backs that with a wide margin on long context (84.2% on GraphWalks 256K to 1M, against 71.8% for Astra) and its DeepSWE lead. Google used Argon internally on C/C++ to Rust migrations of up to 800K-plus lines, and still audits and reviews those rewrites before production. Scope the Task so a person can review the result. A million-token run is a review problem before it is a speed win.

Keep the Task on Flash or another Runtime when it is terminal-heavy (Opus 5.5 leads Terminal-bench 4.0 by nine points), looks like FrontierSWE (Astra leads by more than ten), drives a GUI (Astra leads OSWorld-2.0), or is a routine fix. Argon costs about 2.7 times Flash’s per-token price through November 30, and more than five times in December, because Argon’s price doubles a month before Flash’s does. A Claude Code or Codex Agent on the same board is one reassignment away. Matching models to tasks shows how to write that rule down so the next person routes the same way.

Neither rule is final until Argon runs on your own repository.

Keep the setup, swap the model

The pattern underneath is provider neutrality. Frontier models keep arriving with different winners, and a team that rebuilds its workflow for each one rebuilds it every month. Keep the assignment, context, progress, blockers, results, and human review in one place, and the model becomes a value on a Computer. Agents research, execute, test, and report. People set direction, grant authority, and accept the result.

For several Gemini runs side by side, one worktree per Task works at small scale; our parallel Gemini CLI walkthrough shows the pattern. You can build this manually with worktrees, or use Sharkly to manage the workflow.