By Route Key EditorialPublished 7 min read

Gemini 4 Argon API: Pricing, Long Decode, and Migration

Gemini 4 ArgonGoogle AIAPI migrationAPI pricing

Gemini 4 Argon API pricing, long decode, and migration with RouteKey branding

Is Gemini 4 Argon available through an API?

Yes. Google opened Gemini 4 Argon to developers on October 13, 2026, through the paid Gemini API. The documented model ID is gemini-4-argon, and Google's current model page labels it stable. Early articles describing API access and input limits as unconfirmed reflect the September announcement period.

The practical change is a shared 1,048,576-token window for input, thinking, and output, plus long-decode support that can continue generation across multiple requests. This creates more room for difficult coding and research work, but it also changes how applications should handle output limits, continuations, caching, and timeouts.

The API facts and prices below were checked on October 29, 2026. Benchmark figures are attributed to Google's launch report; RouteKey has not run an independent Argon evaluation.

What changed between announcement and developer release?

Google's September 30 announcement focused on long-running software engineering, knowledge work, and access for trusted cybersecurity defenders. It announced a large output allowance while broader access was still pending.

The October 13 developer release and subsequent model documentation now establish a more concrete implementation contract:

AreaCurrent documented behavior
API modelgemini-4-argon, stable, paid tier
InputText, images, video, audio, and PDF
OutputText
Context1,048,576 tokens shared by input, thinking, and output
Reasoning controllow, medium, or high; medium is the default
CachingImplicit caching; explicit cache creation is unsupported
ToolsFunction calling, code execution, search grounding, File Search, and other documented tools
Computer useSupported in preview

Audio and image input do not imply native audio or image generation. The same model page explicitly marks audio generation, image generation, and the Live API unsupported.

How the shared 1M window actually works

An application cannot supply a million tokens of input and then expect another million tokens of output. All three categories compete for the same window.

For example, with 300,000 input tokens, the theoretical remaining capacity is 748,576 tokens for thinking and output together. A requested output allowance may impose a smaller ceiling. Thinking also consumes that allowance, so the visible answer can be much shorter than max_output_tokens.

This matters for codebase agents. Sending every file, every historical tool result, and every earlier plan can leave less space for the reasoning needed to finish the job. Preserve relevant context and compact stale execution details even when a large window is available.

Google's summary model page lists 65,536 as the default synchronous output allowance. Its more detailed generateContent migration guide distinguishes the underlying transport limits:

Calling patternDocumented per-request or per-stream cap
Non-streaming REST request32,768 tokens
Streaming request65,536 tokens
Long decodeCan continue beyond those caps, within the shared context and configured output budget

The SDK can hide continuation requests behind one method call, which explains why a single "synchronous limit" is not enough to configure a client. Follow the guide for the API surface your application actually uses.

Long decode requires continuation-aware clients

Google documents automatic continuation in Python google-genai 2.30.1 or later and JavaScript @google/genai 2.29.1 or later. When the output budget exceeds the per-request cap, those SDKs can continue and combine the result.

REST clients, older SDKs, and third-party adapters need their own supported continuation behavior. A response can contain no visible text because a whole request segment was spent thinking, while still returning a continuation_token. Treating that as an empty-answer failure could restart expensive work unnecessarily.

A useful implementation review covers four cases:

  1. A normal response finishes within one request.
  2. A thinking-only segment returns a continuation token.
  3. The generation reaches the configured maximum and needs a larger budget to resume.
  4. The user cancels or the network disconnects during a long run.

Google's guide recommends allowing at least 20 minutes per HTTP call for these workloads. That does not mean a person should stare at an inactive chat box for 20 minutes. Long jobs need visible progress, cancellation, and an application-level deadline. Interactive requests should retain a smaller, deliberate latency budget.

What does Gemini 4 Argon cost?

Google's standard paid-tier pricing gives the following schedule. All prices are USD per million tokens and are direct Google prices, not a RouteKey quote.

Token categoryThrough November 30, 2026From December 1, 2026
Input$2.00$4.00
Cached input$0.10$0.20
Output, including thinking$10.00$20.00

The model is unavailable on the free tier. Search grounding and other tools have separate billing rules, and Google counts internal tool instructions and function declarations when processing grounded requests.

Consider 100,000 uncached input tokens and 50,000 billable thinking/output tokens. The token subtotal is $0.70 at introductory rates: $0.20 for input plus $0.50 for output. At the scheduled December rates, the same usage is $1.40.

This is an arithmetic example, not an estimate of how many tokens Argon needs for a real task. Budget both price periods when deciding whether to adopt it; a workflow that is attractive only during the introductory period may need a different routing policy in December.

What the benchmarks say about Gemini 4 versus Gemini 3

Google's September announcement reports these Argon results:

BenchmarkGoogle-reported Argon resultRelevant workload
DeepSWE v1.177.9%Long-running repository engineering
AutomationBench51.3%End-to-end business workflows
LVBench91.7%Long-video understanding
CWE-bench v168.0%Security vulnerability remediation

These are useful reasons to test the model in those domains. They do not establish an improvement percentage over Gemini 3 Pro unless the earlier model is evaluated on the same benchmark version, tools, and budget.

The distinction is especially important for coding. Comparing a Gemini 3 Terminal-Bench 2.0 score with a Gemini 4 Terminal-Bench 4.0 score measures changes in both the model and the benchmark. Likewise, a headline describing a successful internal code migration does not tell you the number of failed runs, reviewer hours, or billed tokens.

For a Gemini 3 migration, use retained traces and acceptance criteria from your own application. Compare grounded answer quality, accepted patches, tool errors, total latency, and cost. If the old endpoint is retired, use the retained baseline and an available production candidate rather than assuming you can rerun that exact historical model.

Five migration changes worth making explicitly

Replace old sampling and thinking controls

Use thinking_level for Argon. Google says the older thinking_budget maps to the nearest level and produces a warning. Its guide also says temperature, top_p, top_k, and candidate_count are deprecated for this model and ignored if supplied. An old configuration file can therefore appear valid while no longer controlling generation as intended.

Remove explicit cache creation

Calls to create or pass explicit cached content are unsupported. Argon uses implicit caching, enabled for prompts of at least 4,096 tokens, with a 95% input discount on cache hits. Keep stable instructions and reusable context consistent, then inspect usage to see whether they actually hit.

Preserve conversation state through tool calls

Google instructs clients to preserve the full conversation history, including thought signatures. An adapter that extracts only visible assistant text can lose state required by a later tool turn. Test a complete tool round trip before testing a long autonomous job.

Set a generation budget and handle continuation

Choose max_output_tokens with both reasoning and the final answer in mind. Test termination, resumption, and cancellation. Increasing the ceiling creates capacity; it does not guarantee a better answer or limit wall-clock duration.

Validate the actual endpoint

A gateway or coding client may expose a subset of Google's native controls. Check the RouteKey catalog for available models and endpoints, then follow the integration guide. Verify long decode and state preservation on that route before relying on Google's SDK behavior.

FAQ

Is Gemini 4 Argon still in restricted preview?

The current Gemini model page labels gemini-4-argon stable, and Google announced paid developer access on October 13. Individual tools can have their own status; computer use is marked preview.

Does it support one million output tokens in one ordinary request?

Long decode can approach the remaining shared-window capacity across continuation requests. Non-streaming REST and streaming calls have smaller individual caps. Input and thinking reduce the capacity available for the final answer.

Can I reuse Gemini 3 explicit caches?

No. Google's Argon guide says explicit cached content is unsupported. Migrate to implicit caching and verify cache hits through the response usage data.

Should Argon replace every Flash request?

Evaluate it first on difficult coding, research, and tool workflows where longer reasoning may improve completion. Keep a lower-cost route for tasks that already meet your quality target, and compare complete workflow costs before changing the default.

Sources and further reading