Model Access Guide
Gemini 4 Argon: access, pricing and agent setup
Google announced Gemini 4 Argon on September 30, 2026. Here is what the launch means for Hermes Agent and OpenClaw users, what the new benchmark numbers measure, and what you can configure today.
What Google has announced
Google positions Gemini 4 Argon for complex software engineering, professional knowledge work and cybersecurity. Its initial rollout is controlled, with broader paid API and Google AI Ultra access planned later. No fixed public-release date is given.
For an agent operator, the immediate question is access. A high benchmark score does not tell you whether your key can invoke the model, whether your framework supports its endpoint, or whether a long tool loop fits your budget.
How to read the benchmark post
Artificial Analysis lists Gemini 4 Argon (High) with an Intelligence Index score of 53 and a weighted cost of $1.99 per Intelligence Index task, as checked on October 2. That cost is specific to the benchmark workload; it is not an API token rate or the price of every task you give a bot. Rankings can change as models and evaluations are updated.
Google reports 77.9% on DeepSWE v1.1, 51.3% on AutomationBench and 91.7% on LVBench in its launch announcement. These are vendor-reported results. Use them to identify workloads worth testing once access is available, then compare success rate, tool errors, latency and total cost on your own tasks.
Pricing and the two different million-token claims
| Item | Announced or reported value | Source / qualification |
|---|---|---|
| Introductory input / output | $2 / $10 per 1M tokens | Google launch announcement |
| Cached input | 95% discount; $0.10 per 1M at introductory input rate | Calculated from Google's announced discount |
| Later input / output | $4 / $20 per 1M tokens | After introductory period; no end date announced |
| Maximum output | 1M tokens | Google announcement; an output limit |
| Context window | 1M tokens | Artificial Analysis model listing; separate from output |
For example, 100,000 uncached input tokens plus 20,000 output tokens calculate to $0.40 at the introductory rates, or $0.80 at the announced later rates. This is token arithmetic, not a quote for an agent session: add all calls and retries, apply the provider's actual billing rules, and budget hosting separately.
A maximum output length is a ceiling, not a recommended response size. Keep agent output bounded to the task. Also check endpoint-specific limits: the context number and output number should not be added together to assume a two-million-token request.
Who can get access?
The Fairwind programme accepts applications from eligible cybersecurity organisations, including governments, critical infrastructure and core technology providers. Approved access is for cybersecurity work and carries restrictions on sharing, redistributing or selling access. Applying does not guarantee approval.
A normal Google API key, a consumer Gemini plan, or an OpenClaw Launch subscription is not proof of Argon access. Do not buy hosting solely on the assumption that Argon is already selectable. Check your provider's entitlement, supported API and model identifier first.
Hermes Agent: configure an available Gemini model
Hermes already has a native Google Gemini provider. On a self-hosted installation, put your own GEMINI_API_KEY or GOOGLE_API_KEY in ~/.hermes/.env, then run:
hermes model
# Choose More providers → Google AI Studio
# Select a model available to your keyThe following ~/.hermes/config.yaml example selects Gemini 3.8 Flash, not Argon:
model:
provider: gemini
default: gemini-3.8-flash
base_url: https://generativelanguage.googleapis.com/v1betaHermes uses the native model ID with provider gemini. Start a new session to evaluate the selected model. See the Hermes Gemini guide for the full setup.
OpenClaw: use Google's current catalogue
For self-hosted OpenClaw, the Google provider documentation verifies API-key onboarding and model discovery:
openclaw onboard --auth-choice gemini-api-key
openclaw models list --provider googleIf your account lists Gemini 3.8 Flash, merge this model setting into your existing configuration. It is an available-model fallback, not an Argon configuration:
{
"agents": {
"defaults": {
"model": { "primary": "google/gemini-3.8-flash" }
}
}
}OpenClaw requires the google/ provider prefix. Keep your own Google key available to the running gateway; its documented environment file is ~/.openclaw/.env. Preserve the rest of your configuration when changing the model.
On OpenClaw Launch, and when Argon becomes available
For a managed Hermes Agent or OpenClaw instance, connect your own Google key and choose an accessible model from the instance card's model menu. Missing Argon access is not something a server restart or a new key name can fix.
- Confirm that the provider explicitly grants your account Argon access for your intended use.
- Read the provider's exact model ID, endpoint, pricing, tool-calling support and token limits.
- Check your framework's current provider support before changing the model.
- Evaluate a small, bounded task before moving an existing automation. Compare cost and completion quality with your current model.
This page will need a verified public API route before it can offer an Argon setup command. In the meantime, Google's public Gemini catalogue and the two framework guides above are the practical starting points.
Sources and update date
Checked October 2, 2026: Google launch announcement, Fairwind access programme, Artificial Analysis model listing, Google API model catalogue, Hermes Gemini documentation and OpenClaw Google documentation.
Gemini 4 Argon FAQ
Can I use Gemini 4 Argon on Hermes Agent or OpenClaw now?
General access is not verified as of October 2, 2026. Google is rolling Argon out to approved Fairwind partners. Both frameworks support Google Gemini, but that does not grant Argon access. Use an available Gemini model until your provider explicitly supplies Argon access and its API documentation.
What is the Gemini 4 Argon API model ID?
We have not verified a public API ID. Do not guess one from the model name or paste an invented OpenRouter slug into your configuration. Check the Google model catalogue and your provider's documentation when general access arrives.
How much does Gemini 4 Argon cost?
Google announces introductory rates of $2 per million input tokens and $10 per million output tokens, with a 95% cached-input discount. That implies $0.10 per million cached input tokens. After the introductory period, announced input/output rates are $4/$20. The announcement gives no end date for introductory pricing.
Does one million tokens mean input context or output?
Google's announcement describes a one-million-token output limit. Artificial Analysis separately lists a one-million-token context window. These describe different limits; confirm the actual endpoint's input, output and combined limits before sizing an agent session.
Does a Gemini subscription or Google API key unlock Argon?
Neither is evidence of current Argon entitlement. Google describes later availability for paid API customers and Google AI Ultra subscribers, without a fixed date in the announcement. Fairwind approval is a separate controlled-access process.