- Book: AI Agents Pocket Guide
- Also by me: Thinking in Go (2-book series) — Complete Guide to Go Programming + Hexagonal Architecture in Go
- My project: Hermes IDE | GitHub — an IDE for developers who ship with Claude Code and other AI coding tools
- Me: xgabriel.com | GitHub
On October 13, Google opened Gemini 4 Argon to every paying Gemini API customer. The price on the pricing page is $2 per million input tokens and $10 per million output. Next to it is an end date. From December 1, the same model costs $4 and $20.
If you wire Argon into a product this week, your cost model has seven weeks of shelf life.
What shipped on October 13
Google first showed Argon on September 30, but only to vetted security teams in its Fairwind Program (VentureBeat). On October 13 that changed. Google's announcement makes it globally available to developers and enterprises through the API in AI Studio and the Gemini platform. It is also in the Gemini app for AI Pro and Ultra subscribers, and "for Google AI Ultra subscribers in Google Antigravity," Google's agentic development platform. Reuters frames the launch as Google's attempt to regain ground on OpenAI and Anthropic after months of delays.
The model ID is gemini-4-argon, listed as stable on the models page. Its model page lists a unified 1,048,576-token context window and three thinking levels: low, medium and high. Input can be text, image, video, audio or PDF; output is text. Function calling, structured outputs, code execution, search grounding and computer use (in preview) are supported. There is no free tier, per the pricing page.
Same model, two prices: the introductory rate runs through November 30, then it doubles.
The $2/$10 price has an expiry date
The pricing page lists every Argon rate twice, once through November 30 and once from December 1:
| Tier (per 1M tokens) | Until Nov 30 | From Dec 1 |
|---|---|---|
| Standard input | $2.00 | $4.00 |
| Standard output (incl. thinking) | $10.00 | $20.00 |
| Batch or Flex input | $1.00 | $2.00 |
| Batch or Flex output | $5.00 | $10.00 |
| Cached input (implicit) | $0.10 | $0.20 |
At the launch price, Argon matches Claude Sonnet 5.5 and GPT-6.1 Sol, both at $2/$10 in VentureBeat's comparison. On December 1 it moves to $4/$20, the same price as Claude Opus 5.5 and Meta's Muse Wonder in that table. GPT-6 Astra sits at $10/$50.
Note the "incl. thinking" on the output row. Thinking tokens bill as output, so a high thinking level adds cost to every call, short answers included.
Price your feature at the December rate now. If the margin only works at $2/$10, your December invoice is where you find out.
Where Argon ranks outside Google's own charts
Google's launch post leads with results Google or its partners reported. Two independent indexes give a second view.
On the Artificial Analysis Intelligence Index, Argon (high) scores 53 and ranks #8 of 216 models. Per VentureBeat, that ties GPT-6 Astra and Claude Fable 5.1, and trails Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56). On the Vals Index as of October 12, the order flips: Argon leads with 68.90%, ahead of Sonnet 5.5 at 67.04% and Opus 5.5 at 66.97%, the same report says.
Google's own coding numbers are split too, and VentureBeat notes they have not been independently replicated. Argon leads DeepSWE v1.1 at 77.9% against Opus 5.5 at 74.2%. On Terminal-Bench 4.0, Opus 5.5 leads, 66.4% to 57.4%.
Sam Witteveen of Red Dragon AI put it this way to VentureBeat: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is." Leaderboards disagree about Argon, so test it on your own tasks before you switch.
The 1M output tokens come with conditions
Google calls it "an industry-leading 1M unified input-output token limit via the Gemini API" (Google). Read the model page before you plan around it. By default, a synchronous request caps output at 65,536 tokens. The 1M figure needs "long decode."
The long decode docs spell out the rules. You go through the Interactions API with background: true and set max_output_tokens above 65,536. Background interactions require store=true. Ask for more than 65,536 tokens without background mode and you get a 400 error. A background run can last up to 24 hours, and you pay for every token in every hop, thinking-only hops included.
The create call, adapted from the Python sample in those docs:
from google import genai
client = genai.Client() # reads GEMINI_API_KEY
interaction = client.interactions.create(
model="gemini-4-argon",
input="Port this C module to Rust: ...",
background=True,
store=True,
generation_config={
"max_output_tokens": 262144,
"thinking_level": "high",
},
)
You then poll the interaction until its status leaves queued or in_progress, and join the text from its output steps. The docs show the polling loop in Python, JavaScript and REST.
Run the arithmetic at list price. A full 1M-token decode is about $10 of output at the launch rate and $20 from December 1, thinking included, plus the input re-sent on every hop (implicit caching discounts the repeats). That is a job for a queue and a budget check, not for a request handler with a 30-second timeout.
Long decode runs in the background: you start the job, then come back for the result.
Before you swap the model ID in your agent
Google says Argon was trained to "pause for human confirmation before taking irreversible steps" (Google). It also reports that Argon leads the Gray Swan indirect prompt injection benchmark, with attack success at 0.7% across 15 adaptive attempts (Google Cloud). Those are Google's numbers. They do not replace your own approval gate on anything that deletes, pays or deploys. If your agent does security work, note that the public model "has cyber mitigations that block offensive tasks such as pentesting and exploitation," per the Google Cloud post.
For coding agents, the fair test is the same ticket you give your current agent. I build Hermes IDE, so weigh this with that in mind. Hermes runs each agent's own CLI in a real terminal, each task on its own hermes/<name> branch and worktree. Antigravity CLI is in its agent list, next to Claude Code and Codex. Start the same task in two of them with ⌘N (Ctrl+Shift+N on Windows and Linux), then open each agent's changes, turn by turn, in the Review Desk (⌘G, or Ctrl+Shift+G) and compare them yourself.
Two caveats. I haven't checked which Google plans see Argon in the CLI's model list, and I haven't run Argon myself. Hermes is free to download and source-available (BSL 1.1). It runs on macOS, Windows and Linux and is not affiliated with Google, Anthropic, OpenAI or Nous Research. Download it or read the code on GitHub.
Will you build on Argon at $2/$10, knowing the bill doubles on December 1? Or wait for the December price and compare it with Opus 5.5 then? Tell me in the comments.
If this was useful
If you are wiring a new model into an agent this week, the AI Agents Pocket Guide covers the loop, tool and approval patterns that keep a model swap from breaking production. And if you run more than one coding agent, Hermes IDE puts them side by side, each on its own branch.
Top comments (0)