JOURNAL / 2026.10.01

Gemini 4 Argon expands long-running work, but reaches approved defenders first

Google introduces its new generation with a million output tokens and competitive results in professional work and cyber defense; restricted initial access and differences between evaluations matter as much as the announced capability.

Gemini 4 now has a name: Argon. Google announced it on September 30 with a focus on professional tasks that require continuing well beyond the first response. Its debut, however, does not let anyone replace an application's model today: it starts with a group of approved cybersecurity defenders in the Fairwind Program. Wider availability will begin with paid API customers and Google AI Ultra subscribers, with no specific date in the announcement.

The clearest technical change is a one-million-token output limit, up from the previous 64,000. That is room for reasoning and generation along a trajectory, not just for receiving an enormous document. Google also reports that Argon agents replaced 32,000 lines of SIMD code in a Rust port of libgav1 with safe Rust that the compiler can vectorize. According to the company, the result preserves video output and runs 2.7 times faster than that port; the comparison is not against the optimized C++ decoder.

That example helps explain the bet: a series of profile-guided experiments and inspections of compiler output, with a check external to the generated text. It remains a provider-described case. It does not show that an arbitrary migration can run unattended without review.

More runway does not guarantee a better landing

A higher limit prevents work from stopping for lack of space. It also allows more time to be spent pursuing a mistaken hypothesis. For someone building agents, the useful change would be finishing a difficult investigation or modification without losing the thread; the relevant test is the verified result, along with time, cost, and errors, rather than trajectory length.

Google's results page offers an encouraging signal and a warning. On DeepSWE 1.1, which evaluates long-horizon software engineering, Argon scores 77.9%, against GPT-6 Astra's 74.1%. On FrontierSWE v2, another coding evaluation, it trails Astra and Opus 5.5. Capability has a shape; winning one row does not establish general superiority.

The methodology document explains why those rows need care. Google ran DeepSWE with the mini-swe harness, while taking rival results from public leaderboards or system cards. In other tests, it ran every model itself. By default, it uses the highest reasoning level and single-attempt results, with documented exceptions. This is not one competition with every system under identical conditions.

How success is measured matters too. OSWorld 2.0 reports partial credit on its offline subset, rather than only completed workflows; Google selects the best of three complete runs. For video, models receive different frame counts because of API limits. These are explicit, auditable choices, but they prevent the whole table from becoming a clean measurement of the isolated model.

There is outside evidence: Artificial Analysis gives Argon at High effort 53 points on its index, composed of ten evaluations of work, knowledge, coding, and reasoning. It records a weighted cost of $1.99 per task and 110 million output tokens across the evaluation. That supports its status as a frontier model; it does not independently validate the internal migrations or the reliability of a run using the entire new limit.

The first product is a relationship of trust

Argon's cyber capability description combines vulnerability discovery in repositories, penetration testing without source code, and patch generation. These are different jobs: finding a flaw, demonstrating it, and repairing it require different evidence.

CWE-bench v1 lets us check part of that promise outside Google's account. Its 120 private tasks provide a repository without naming the flaw and require an audit and fix. Argon reaches 68% on the single-attempt programmatic verifier, tied with Astra and Grok 4.7. The verifier checks that the exploit no longer works and existing tests pass. This is defensive capability evaluated in a bounded environment, not a discovery rate for unknown flaws in any system. Each model also uses a different harness.

The Fairwind Program adds the other half of the product: organization screening, authorized defensive use, individual authentication, phishing-resistant MFA, and employee access tracking. It prohibits redistributing or selling access. Its page describes more than 650 partners, but specifies that only a subset receives Argon; the program's size must not be confused with this model's deployment.

Google says the version for defenders and its own internal teams will be offered without the model's cyber guardrails. Fairwind retains organizational obligations. Removing a model filter does not remove authorization limits: it shifts part of control toward who gets in, what they may do, and how use is recorded.

My reading is that Argon deserves attention for this combination more than for counting benchmark victories. It expands the room for sustained work and introduces it first where teams can verify findings. That is a defensible strategy, although it delays broader testing and leaves the provider selecting much of the initial evidence.

For a team without access, planning changes today, rather than the endpoint. It makes sense to prepare representative tasks with acceptance criteria and a budget, ready to evaluate the version that eventually arrives. The announced rate also has two phases: $2 per million input tokens and $10 per million output tokens during the introductory period; afterward, $4 and $20. The period's duration is not specified.

The question that will decide Argon's value is concrete: can it turn more runway into more verified work, without spending and supervision growing at the same pace? A million tokens creates room to answer it. It is not yet the answer.

Sources

Documentation and results accessed on October 1, 2026.

← Back to journal