Gemini 4 Argon was announced on 30 September 2026. Its first rollout is to trusted testers and cyber defenders through Google's Fairwind Program. Broad consumer access is still ahead. If you saw the launch and immediately opened the Gemini app, that distinction matters. Google announcement.
I would start with the access notice, then the capability claims. An impressive model is useful to you when you can run it on the work you actually need done.
The million-token number is an output limit
Google announced a one-million-token output limit, increased from 64,000. This provides room for longer reasoning and generation. It is not a claim that every answer will be that long, and it should not be relabelled as the input context window. Google positions Argon for coding, knowledge work and defensive cybersecurity. Launch details.
Longer output can help a model stay with a difficult problem. It also creates more material to review. I would judge a completed migration by the tests it passes and the behavior it preserves, rather than the number of tokens it produces.
What Google says it has done
Google reports using Argon for large code migrations and data-center memory optimization, including more than 300 TiB of memory freed after rollout. These are Google's results. I have not independently tested Argon or reproduced those outcomes. Reported internal work.
Treat those examples as reasons to investigate a capability. Your small app, unfamiliar language and review process may produce a very different result.
Announced API pricing
| Period | Input, per million tokens | Output, per million tokens |
|---|---|---|
| Introductory | $2 | $10 |
| After the introductory period | $4 | $20 |
Google also announces a 95% discount on cached input. The launch article does not give an end date for the introductory period. These are announced API prices, not a promise that your account already has access. Broader rollout is planned to start with paid API customers and Google AI Ultra subscribers. Pricing and rollout.
A useful test to prepare now
Choose one bounded task with a known answer or a checkable result. For example, migrate one isolated module while preserving its public interface. Keep the existing implementation and tests as a baseline.
Use this worksheet when you have access:
Task:
Model and access date:
Input files and relevant context:
Baseline result:
Allowed changes:
Actions requiring my approval:
Acceptance checks:
Input, cached input and output tokens:
Elapsed time:
Actual billed cost:
Checks passed and failed:
Human corrections needed:
Decision: use again / investigate / keep baseline
Save failures too. A successful screenshot without the failed attempts tells you very little about whether the workflow is repeatable.
A prompt for your first controlled evaluation
This is an example prompt, not a benchmark I have run:
Inspect this module and its tests. Propose the smallest migration plan.
Preserve the public interface and expected behavior.
List missing inputs before changing code.
Work only in the selected branch and files.
After implementation, run the acceptance checks and report failures.
Do not deploy, contact anyone, or change external services.
Show the diff, test evidence and unresolved questions for review.
Change the boundaries to fit your project. Review a small change before giving the model a larger codebase.
Questions people are asking
Can everyone use Gemini 4 Argon today?
The launch describes restricted initial access, with wider availability planned. Check the official notice and the actual model list in your account before paying for access.
Does a million-token output limit guarantee better answers?
No. It describes available headroom. Your acceptance checks still decide whether the output is useful.
Should I change my whole workflow immediately?
Start with one comparison against your current setup. Keep the input, success criteria and review effort visible. A useful model earns more responsibility through results you can check.
Sources
Checked 1 October 2026. Access and pricing can change. This is a sourced launch explainer with an illustrative testing framework, not a hands-on Argon review.