# Gemini 4 Argon: 1M Context, GA Pricing and What Changes Dec 1

URL: https://www.devopsness.com/blog/gemini-4-argon-1m-context-limited-access
Author: Kiril Urbonas
Published: 2026-10-01
Updated: 2026-11-15
Categories: AI
Tags: Gemini, LLM, AI, Security, AI Agents

Summary: Google shipped a 1M-token output ceiling and strong cyber-defense benchmarks, then handed the keys to almost nobody. Here is what that gating pattern tells you.

---
# Gemini 4 Argon: 1M Context, GA Pricing and What Changes Dec 1

You can use Gemini 4 Argon now, if you pay. Google launched it on September 30, 2026 for vetted Fairwind Program cyber defenders only, then made `gemini-4-argon` generally available on the paid Gemini API tier on October 13. The free tier still gets nothing. Introductory pricing is $2 input and $10 output per million tokens through November 30, 2026.

> [!VERIFIED]
> Gemini 4 Argon (`gemini-4-argon`, stable), against Google's launch and developer-availability posts, the Gemini API model page, changelog and pricing page. Checked October 2026.

We wrote the first version of this post on launch day, when "mostly can't use yet" was literally true. Two weeks later the gate opened for paying API customers, and Google AI Ultra subscribers can try it in AI Studio and Google Antigravity. The gating pattern is still the more interesting story, and it is why we kept the original argument below.

## The headline number is output, not context

Every "1M context window" headline about Argon was half right. Google's launch post led with a 1-million-token **output** ceiling, up from 64,000 tokens on the prior generation. The model page now spells out the mechanics: a single 1,048,576-token context window shared between input and output, a 65,536-token default output limit for synchronous requests, and up to 1,048,576 output tokens with long decode. So "it reads a million tokens" and "it writes a million tokens" both hold, but not in the same request, because they draw from one window.

The distinction matters operationally: a bigger input window changes what you can stuff into a prompt, while a bigger output ceiling changes what a long-horizon agent can produce inside one trajectory, meaning fewer stop-and-reprompt cycles for a full vulnerability remediation pass or a multi-file refactor that previously had to be chunked across calls.

## Why it was gated behind cyber defenders first

Google rolled Argon out first through the Fairwind Program, a pool of trusted cybersecurity teams including Wiz, which uses the model in its Scan for Good initiative to find and fix exposures in public infrastructure. Argon ties for first on CWE-bench v1, a vulnerability-remediation benchmark, at 68%, and Google pitches it explicitly for vulnerability identification, validation, and patching. That is a dual-use capability in the plainest sense: a model good enough to find and fix a zero-day is also good enough to help someone build one.

Shipping it first to defenders while the team kept tightening misuse prevention and injection resistance is Google choosing to let the people with the most to lose from misuse stress-test it first. Google also says it worked with the U.S. government's voluntary pre-release access process. That is a more cautious rollout than a straight waitlist, and it signals Google treating agentic cyber capability as its own risk category, separate from a routine capability bump.

## What the strong benchmarks actually unlock

Google reports Argon at 77.9% on DeepSWE v1.1, a long-horizon software engineering benchmark, and calls it state of the art. It also claims the lead on the Vals Index, Vals Finance Agent v2, Harvey's Legal Agent Benchmark, Zapier's AutomationBench (51.3%) and LVBench for long video (91.7%). Google published no competitor scores alongside those numbers, so treat the rankings as vendor-reported until independent runs exist.

These are exactly the long-horizon, document-heavy, multi-step workloads you'd expect a bigger output ceiling to help with: an agent that drafts a full legal memo, reconciles a quarter of transactions, or walks an entire codebase end to end without truncating mid-task. That is a specific, named set of workloads where generating more tokens per run before re-invoking the model is a real unlock, not a vanity spec.

## The honest caveat: the coding-agent case is unproven

This is the part the launch framing undersells. Google's launch post reports no Terminal-Bench result, the benchmark built around real command-line and agentic coding tasks, and the Terminal-Bench numbers circulating in third-party comparison tables disagree with each other, some putting Claude Opus 5.5 well ahead. We couldn't confirm any of them against a primary source, so we won't repeat them. A model pitched partly on coding and autonomous patching hasn't shown, in its own launch materials, that it beats the agents you already run on CLI-heavy work. Until it does on your eval set, that is not a reason to switch.

## What to actually do about this today

Now that it is callable, the call shape is the standard `google-genai` pattern: one generation call with the long input batched in a single request rather than split across calls.

```python
# argon_report.py
from google import genai

client = genai.Client()  # paid-tier API key; Argon is not on the free tier

response = client.models.generate_content(
    model="gemini-4-argon",
    contents=full_case_file_text,
    config={
        # 65,536 is the synchronous default; going toward 1M needs long decode
        "max_output_tokens": 65_536,
        "system_instruction": "Draft a full remediation report, not a summary.",
    },
)
print(response.text)
```

Put a cost budget on it before you put a retry policy on it: at the introductory $10 per million output tokens, a single full 1M-token generation costs about $10, and that rate doubles to $20 on December 1, 2026. Batch requests are half price. Keep evaluating the providers already in your [best LLM APIs](/blog/best-llm-apis) rotation, and if your workload is agentic and security-sensitive, the controls in our [AI agent security guide](/blog/ai-agent-security) apply regardless of which vendor wins this slot. For a broader build decision, [OpenAI vs Anthropic vs Gemini](/blog/openai-vs-anthropic-vs-gemini-api) still frames the trade-offs.

## The decision, concretely

- **Should we switch our coding agent to Argon?** Not on what's public. Google's launch materials report no Terminal-Bench number, so run it against your own agentic coding evals before moving anything.
- **Should we request Fairwind access if we do security work?** Yes, if you are a cyber-defense team with a real use case. The CWE-bench score and the Wiz deployment suggest genuine capability for vulnerability work, not just a demo.
- **Should we redesign our long-document pipeline around 1M output?** Prototype it now, on Batch pricing, and budget against the December 1 rate rather than the introductory one.
- **Should we care about the gating pattern itself?** Yes. Defenders first, paying API customers two weeks later, free tier not at all, is a template worth watching for the next dual-use release.

## Questions people ask

### Is Gemini 4 Argon available in the API?

Yes, since October 13, 2026, on the paid tier of the Gemini API in Google AI Studio, under the stable model code `gemini-4-argon`. It is not available on the free tier. Google AI Ultra subscribers can also try it in AI Studio and Google Antigravity. Before October 13, access was limited to vetted cyber-defense teams in the Fairwind Program.

### Does Gemini 4 Argon have a 1M context window or 1M output?

Both, sharing one budget. The model page lists a 1,048,576-token context window for input and output combined. Synchronous requests default to 65,536 output tokens, and long decode raises output to as much as 1,048,576 tokens. A request that uses most of the window for input leaves correspondingly less room for output.

### How much does Gemini 4 Argon cost?

Through November 30, 2026, standard pricing is $2 per million input tokens and $10 per million output tokens, including thinking tokens. From December 1 it doubles to $4 and $20. Batch requests cost half the standard rate, and context caching is listed at $0.10 per million tokens during the introductory period.

## The call we'd make

Gemini 4 Argon is a legitimate capability jump on long-horizon output and a genuinely interesting signal about how Google handles dual-use risk, and since October 13 you can put it in a production plan if you're on the paid tier. Default to the providers your evals already trust for agentic coding, prototype Argon on the long-document and remediation workloads its output ceiling actually helps, and budget against the December 1 price, not the launch-week one. The missing Terminal-Bench number should anchor your expectations more than the 1M headline.

## Sources

- [Google: Introducing Gemini 4 Argon (September 30, 2026)](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/)
- [Google: Gemini 4 Argon is now available for developers (October 13, 2026)](https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/)
- [Gemini API Docs: Gemini 4 Argon model page](https://ai.google.dev/gemini-api/docs/models/gemini-4-argon)
- [Gemini API Docs: Pricing](https://ai.google.dev/gemini-api/docs/pricing)

