← All Guides

Model Guide

Gemini 4 Argon: physical reasoning, pricing and agent setup

Argon is publicly rolling out. Learn what the physics and 3D demos show, how to test them, and how to configure Gemini on Hermes Agent or OpenClaw.

Updated October 13, 2026: Google has announced the public rollout. This replaces our October 2 restricted-preview coverage. The setup below follows upstream documentation; we have not run an Argon inference or reproduced the media demos ourselves.

What is Gemini 4 Argon, and who can use it?

Google's October 13 release expands Argon beyond the September Fairwind preview. It targets reasoning, software engineering and professional work. Fairwind remains a separate cybersecurity programme; it is no longer the only access path.

SurfacePublished access route
Gemini app and AI Mode in SearchGoogle AI Pro or Ultra; account availability and quotas apply
Google AntigravityGoogle AI Ultra
Gemini APIPaid API tier; separate API billing
AI Studio model experienceUltra subscriber access; developers can build through the paid API tier

A model appearing in a consumer app does not establish which API your bot can call. Confirm the key's access, the provider route and your agent version.

What does “physical reasoning” mean?

Google's robotics post reports 87% on ERQA, an embodied spatial reasoning benchmark. Its examples connect visual observations to simulation code and robot actions, including sketch-to-MuJoCo reconstruction and a simulated twelve-ball weighing task. These are vendor-reported results under specific tooling and conditions, not proof of reliable control in every environment.

The useful pattern for an agent is to turn observations into a model, execute a bounded experiment, and check the result. An attractive animation alone does not establish correct mechanics. Geometry, units, constraints and solver outputs need separate checks.

Reading the article's 3D and video tests

The Chinese hands-on report you shared describes a Fourier-series pelican animation, a Yingxian wooden pagoda reconstruction, Titanic clips and scientific visualisations. Treat these as examples to investigate, not our own test results. Its pagoda test notes a floor-count mismatch, and its video tests describe short clips and awkward Chinese pronunciation.

Argon's model specification lists multimodal inputs and text output. Argon can produce code, plans or tool calls that drive Blender, a browser renderer, or separate video and speech models. A narrated clip from a combined workflow does not mean the Argon API directly generates video or voice.

Reported speed and cost comparisons vary with the agent harness, reasoning effort, caching and retries. Compare identical tasks with the same assets and acceptance criteria; record failed attempts as well as the best output.

Three practical tests you can repeat

These are original evaluation prompts, not measured Argon results. Use a disposable workspace with a browser or renderer, save the generated source, and ask for evidence of execution.

1. Fourier animation: test the mathematics

Create a single HTML canvas demo approximating a closed bicycle-and-bird
outline with discrete Fourier coefficients and rotating epicycles.
Include 10, 30 and 100 harmonic settings, pause and reset controls.
Explain sampling, coefficient calculation and the reconstruction equation.
Show reconstruction error at each setting. Use no external assets.

Check that changing the harmonic count changes the reconstructed path, pause actually stops time, and the explanation matches the implementation. Decorative circles without Fourier coefficients do not pass.

2. Reference-image 3D: test structure fidelity

Using the attached reference photo, build an interactive pagoda model.
First list visible storeys, roof tiers, symmetry and uncertain dimensions.
Preserve those counts. Add orbit controls and labelled structural parts.
List assumptions and compare the final model with the reference.
Save source and a screenshot; report what you could not reproduce.

Count storeys and roofs independently, compare silhouettes, inspect hidden sides, and test camera controls. Plausible architecture is a different criterion from matching the reference.

3. Physics simulation: test numerical behaviour

Build a projectile-motion demo with SI units, adjustable launch speed
and angle, gravity and a fixed time step. Show position and velocity.
Compare the no-drag trajectory and landing time with analytic equations.
Add reset and pause controls. Report maximum numerical error and
explain whether any visual effect is decorative rather than simulated.

Compare the no-drag result with a known analytic solution, then halve the time step and check convergence. A research-grade fusion solver needs domain validation beyond a browser demo.

API pricing and token limits

The following are Google standard API rates, checked October 13. Hosting, aggregator routes and separate media tools are billed according to their own terms.

Per million tokensThrough Nov 30, 2026From Dec 1, 2026
Uncached input$2$4
Output, including thinking$10$20
Cached input on eligible implicit cache hits$0.10$0.20

For 100,000 uncached input tokens and 20,000 billable output tokens including thinking, token charges calculate to $0.40 during the introductory period or $0.80 afterward. Sum every model call, retry and tool charge for the full task.

The documented window is 1,048,576 tokens shared by input and output, with 65,536 default synchronous output tokens. Larger generations need long decode and continuation, and input reduces the remaining output room. Continuation rereads prior tokens, so billed input can exceed the unique context size.

Google's migration guide documents low, medium and high thinking, with medium as the default. Argon supports implicit caching, not explicit cache objects. Verify these behaviours in your installed adapter before relying on long decode.

Self-hosted Hermes Agent setup

Current upstream Hermes documentation lists Argon on the native gemini provider. Store your own GEMINI_API_KEY or GOOGLE_API_KEY in ~/.hermes/.env. Run hermes model and choose Google AI Studio, or merge this mapping into ~/.hermes/config.yaml:

model:
  provider: gemini
  default: gemini-4-argon
  base_url: https://generativelanguage.googleapis.com/v1beta

Start a new session. Use the native ID without a google/ prefix. The upstream adapter documents continuation support, but an older packaged version may lack it. Follow the Hermes Gemini setup guide and check your installed version if tool calls or long replies fail.

Self-hosted OpenClaw setup

The upstream Google provider supports API-key onboarding and model discovery. Configure your own key in ~/.openclaw/.env for the gateway process, then inspect your catalogue:

openclaw onboard --auth-choice gemini-api-key
openclaw models list --provider google

If Argon is available to your key and installation, merge this model selection into ~/.openclaw/openclaw.json, preserving your other settings:

{
  "agents": {
    "defaults": {
      "model": { "primary": "google/gemini-4-argon" }
    }
  }
}

OpenClaw uses a provider-prefixed ID. Model discovery alone does not verify a complete tool loop or long decode. Use the OpenClaw Gemini guide for provider setup and assess those behaviours before migrating an automation.

Managed hosting: check the actual model menu

Hermes and OpenClaw instances on this site share the model selector. In Dashboard, open the bot's model menu, search for Argon and inspect the billing route. A Google-key route, an aggregator key and platform credits have different credentials and costs.

The managed catalogue can be smaller than upstream. We have not verified an Argon inference on a managed instance, so this guide does not promise that every bot can select it today. If it is missing, retain a working model; editing an ID does not establish compatibility or entitlement.

  1. Record the model, provider, reasoning setting and installed agent version.
  2. Use a small task with a checkable answer, then one bounded tool task.
  3. Compare quality, elapsed time and total billable tokens across several runs.
  4. Keep the previous model settings available before moving existing automations.

Sources and update date

Updated October 13, 2026 using Google's public release, robotics demonstrations, model specification, API pricing, migration guidance, Hermes upstream and OpenClaw upstream. The supplied Chinese report is secondary commentary; its snapshot displays an October 14 Beijing publication date.

Gemini 4 Argon FAQ

Can I use Gemini 4 Argon on Hermes Agent?

Current upstream Hermes documentation lists Argon under its native Gemini provider. Use your own paid-tier Google API key and model ID gemini-4-argon. A managed installation may run an older version or expose a smaller catalogue; check its model menu before switching.

Does Google AI Pro include the Argon API or Antigravity?

Google lists Pro and Ultra access in the Gemini app and AI Mode in Search. Antigravity access is listed for Ultra. API calls require separate paid-tier API access; a consumer subscription does not supply API billing credits.

Can Gemini 4 Argon generate video, images or speech?

The Argon API accepts text, images, video, audio and PDF, and outputs text. It can write rendering code and tool instructions. Finished media needs rendering or separate generation tools; it is not a native Argon API output.

How much does Gemini 4 Argon cost?

Google standard API introductory rates are $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1 they become $4 and $20. Thinking tokens count as output. Check your chosen provider for its current rates and additional tool charges.

Can I use a full million tokens for both input and output?

No. The shared input and output window is 1,048,576 tokens. The documented default synchronous output limit is 65,536; longer generation needs the continuation protocol. Input occupies part of the same window.

Related guides

Choose a Gemini model for your Hermes Agent

Explore managed Hermes hosting. Check the model catalogue and billing route before choosing Argon; API usage and media tools have their own costs.

Explore Hermes hosting