Gemini 4 Argon is available. Budget beyond the launch price.

Gemini 4 Argon is now generally available through the Gemini API. Google records the release on 13 October 2026, following its 30 September announcement and initial rollout to selected cyber defenders. Those are different milestones: the earlier announcement did not establish broad developer access. API release notes, original announcement.

For a business considering a pilot, the useful questions are what it can access, what changes in an existing application, and whether the economics still work after introductory pricing ends. This briefing reviews Google's published documentation; Mimir has not benchmarked Argon.

Match access to the product you intend to use

Google's developer announcement directs developers to the paid Gemini API tier in AI Studio and Android Studio. It also identifies Google AI Ultra access in AI Studio and Antigravity. Developer availability.

Consumer access is described separately: Google's launch post includes Pro and Ultra subscribers in the Gemini app and AI Mode. Do not assume that access inside an app provides the same entitlement or billing arrangement as an API integration. Launch availability.

The release material describes hosted services, not downloadable Argon weights. If keeping inference on your own machine is the requirement, this announcement does not establish that option. Our local AI guide for Macs covers a different deployment path.

Price the pilot and the ongoing service separately

Google lists these Standard paid-tier API rates in USD per million tokens, checked on 14 October 2026:

These are published rates, not measured cost per completed task. Batch, Flex and Priority have separate schedules; the chart compares Standard rates only. Tool charges and the amount of reasoning generated can also affect the bill. Google's pricing table.

For illustration, 100,000 uncached input tokens plus 20,000 billable output tokens would cost $0.40 at introductory rates and $0.80 from December, before other charges. This is arithmetic using a hypothetical workload, not a prediction of your usage.

Our advice: record tokens, retries and human review during the pilot, then recalculate the same workload at December's rates before committing to a budget.

A million-token window is a shared allowance

The model specification lists a unified context limit of 1,048,576 tokens. Input, thinking and output share that space; it is not a million input tokens plus a separate million-token answer. The default synchronous output limit is 65,536 tokens. Model specification.

For longer responses through the Interactions API, Google's guide requires background execution and an output limit above 65,536. The service continues generation in successive parts, which the application must collect correctly. Simply turning on background execution without raising the limit leaves the default in place. Long-decode documentation.

Our recommendation is to test completion, cancellation and recovery from interruptions before offering long-running jobs to staff. A larger allowance is useful only when the application can manage the resulting work.

Review the integration before switching models

Google's migration guide identifies changes worth checking in an existing application:

• Use the model ID gemini-4-argon.
• Test the available thinking levels: low, medium and high; medium is the default.
• Review output limits because thinking consumes part of the allowance.
• Remove explicit cache creation and references; this model supports implicit caching instead.
• Do not rely on temperature, top_p or top_k: the guide says they are ignored.
• Allow for longer-running requests and review client timeouts.

These are documented behaviours, not findings from a Mimir migration. Migration guidance.

Our advice is to compare the new configuration against your current one using the same representative tasks. Keep a working fallback until you understand differences in answers, tool calls and cost.

Use the evaluations to choose a test, not promise an outcome

Google's evaluation methodology mixes its own runs with external leaderboards and other providers' reported results. It also documents differences in test setups. A combined headline ranking should not be treated as proof that every model was measured under identical conditions. Evaluation methodology.

The model card acknowledges hallucinations, occasional slowness and timeouts. Those limitations matter for document review and unattended workflows even when benchmark results look strong. Model card.

Start with one bounded task, define what a correct result looks like and retain human approval for consequential actions. Measure time to an accepted result, including corrections, rather than response length or a convincing demonstration. Our guide to reading AI benchmarks explains that distinction.

Argon's wider availability makes a controlled evaluation possible. The decision to adopt it should rest on your task results and sustainable operating cost.