Gemini 4 Argon: A Developer's Guide to Google's 1M-Token Frontier Model

Google announced Gemini 4 Argon on September 30, 2026, and the framing in the launch post, written by Koray Kavukcuoglu, SVP of Google DeepMind and Google Chief AI Architect, is unusually concrete. Argon is built to sustain deep reasoning across complex, long-horizon workflows, and it is rolling out first to a set of trusted cyber defenders through the Fairwind Program before reaching developers, enterprises and consumers. The model covers software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense, with an industry-leading 1 million token output limit. Here is what developers should take from the launch, and what to do while they wait for access.

Gemini 4 Argon is Google's next-era frontier model

What Google announced

According to the Google blog post, Gemini 4 Argon is a new frontier model rolling out to trusted cyber defenders via the Fairwind Program. Google says it is engaged in the U.S. government voluntary process for pre-release model access while it gradually expands availability. Pricing starts at an introductory $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95 percent off the input token price. Google says availability will follow, starting with paid API customers and Google AI Ultra subscribers, after the model is made available to developers, enterprises and consumers as soon as possible.

# Gemini 4 Argon, in one object. Source: the Google blog post by
# Koray Kavukcuoglu (Sep 30, 2026) and TechCrunch coverage.

GEMINI_4_ARGON = {
    "announced": "2026-09-30",
    "first_release": "trusted cyber defenders via the Fairwind Program",
    "input_per_million_usd": 2.00,
    "output_per_million_usd": 10.00,
    "cached_input": "95% off input token price",
    "max_output_tokens": 1_000_000,   # up from 64K
    "focus": ["software engineering", "enterprise knowledge work", "cyber defense"],
    "then": "paid API customers and Google AI Ultra subscribers",
}

# Google is expanding access gradually, with the broader release to follow.

The 1M-token output headline

The capability Google leads with is output length. Argon significantly expands the model output token limit to an industry-leading 1 million tokens, up from the previous 64K tokens. That is roughly a fifteen-fold increase in how much a single trajectory can produce. For agents, the practical consequence is that a single call can emit an entire migration plan, a long patch series, or a full research memo instead of being chopped into fragments and re-stuffed into context. When a model has the headroom to think deeply and generate hundreds of thousands of tokens in one go, the shape of long-horizon tasks changes: more of the plan lives in one coherent context.

# The headline capability is depth, not just size. 1M output
# tokens means one trajectory can carry a long, multi-step plan.

def reasoning_budget(argon):
    return {
        "previous_output_limit": 64_000,
        "argon_output_limit": 1_000_000,
        "gain_x": 1_000_000 / 64_000,   # ~15.6x
    }

# For agents this matters: a single call can emit an entire migration
# plan or patch series instead of being chunked into fragments.

The output limit jumps from 64K to 1M tokens in a single trajectory

Benchmarks that matter

Google reports several sets of numbers, and the ones worth noting span coding, knowledge work and security. Argon sets a new state of the art on DeepSWE v1.1 at 77.9 percent, which measures long-horizon software engineering in realistic conditions. It leads the Vals Index, which weights finance, coding, legal and tax work by their contribution to U.S. GDP, and shows strong results on Vals Finance Agent v2 and the Harvey Legal Agent Benchmark. On AutomationBench, Zapier benchmark for end-to-end business execution, Argon ranks first at 51.3 percent. For long video understanding on LVBench it reaches 91.7 percent, and on CWE-bench v1, which measures vulnerability remediation, it ties for first at 68 percent.

# Calling Argon once it reaches your tier. The API surface follows
# the Gemini family, so migration is a model-name change for most.

from google import genai

client = genai.Client()
resp = client.models.generate_content(
    model="gemini-4-argon",
    contents="Analyse this long-horizon task and return a full plan.",
)
print(resp.text)

# Enterprise knowledge work (finance, legal) and cyber defense are
# the domains Google highlights; start with tasks that fit those.

What Google did with Argon internally

The most persuasive evidence is what Google claims it already did with the model in production. On quantum algorithmic optimization, Argon beat a published baseline by 40 percent in a matter of minutes. A fleet of Argon agents analysed profiling telemetry and applied memory optimisations that freed over 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings. On large-scale language migrations, Argon agents are moving C and C++ codebases to Rust across Google, scaling from libraries like re2 and libgav1 up to the Fuchsia Zircon kernel at more than 800 thousand lines. For libgav1, agents replaced 32 thousand lines of SIMD code and produced a memory-safe decoder that runs 2.7 times faster than the Rust port, with identical output.

# What Google itself did with Argon, from the launch post.

INTERNAL_WINS = [
    "quantum optimisation: beat the published baseline by 40% in minutes",
    "memory efficiency: freed 300 TiB fleet-wide; est. 500 TiB to 1 PiB total",
    "C/C++ to Rust migration: Fuchsia Zircon kernel, 800K+ lines",
    "libgav1: replaced 32K lines of SIMD; memory-safe decoder 2.7x faster",
]

# Every one of these is audited before production, Google stresses.
# Treat them as targets for your own long-horizon agent tasks.

Argon is rolling out to cyber defenders first, then paid API customers

The safety posture

Google is explicit that this is a phased release. The blog describes strengthening safeguards across four areas: defending against misuse such as cyber and CBRN attacks, defending against prompt injection, monitoring for misalignment via chain-of-thought and action monitoring, and hardening sandboxed systems before high-risk training or evaluations. Argon is described as Google most resilient model yet against indirect prompt injections, leading on the Gray Swan benchmark. Notably, Google says it will release Argon without cyber guardrails to trusted defenders and its own internal teams so they can use its full defensive capability, which is why general availability is staged rather than immediate.

# Safety posture, and a checklist for adoption.

ARGON_SAFETY = {
    "guardrails": ["misuse (cyber/CBRN)", "prompt injection", "misalignment", "hardened sandboxes"],
    "prompt_injection": "leading on Gray Swan Indirect Prompt Injection benchmark",
    "cyber_defenders": "released without cyber guardrails to trusted defenders",
}

ADOPTION_CHECKLIST = [
    "confirm your tier: Fairwind, paid API or AI Ultra",
    "budget: $2/M input, $10/M output, cached input 95% off",
    "exploit the 1M output limit for long-horizon tasks",
    "add your own prompt-injection defences for agent tool use",
    "start where Google leads: coding, knowledge work, cyber defense",
]

What developers should do now

Access is staged, so the sensible move is to prepare rather than wait idly. Confirm which tier you will land in, whether Fairwind, paid API or Google AI Ultra. Model the cost of long outputs before you run them, since a million-token trajectory is powerful and billable. Design at least one workflow that exploits the long output limit instead of fragmenting it by habit. If you use Argon for agent tool calls, add your own prompt-injection defences rather than relying solely on the model. And start where Google says Argon leads: real-world software engineering, finance and legal knowledge work, and defensive cybersecurity, then widen from there. Access is likely to stay gated for a while, so the teams that win are the ones preparing workflows and budgets now rather than scrambling when the API opens to everyone.