Gemini 4 Argon: Benchmarks, Access and the Work It Fits

Google announced Gemini 4 Argon on September 30, 2026, positioning it for difficult software engineering, enterprise knowledge work and cybersecurity defense. The published results support a narrower, more useful verdict than “best AI model”: Argon looks particularly strong for document-heavy work and sustained reasoning, but its coding lead depends on the task. Access is also restricted, making this a review of published evidence rather than a hands-on test.^[1]

As of October 6, our verdict is to shortlist Argon for complex research and carefully supervised engineering pilots when access opens, not replace an existing production model on launch-day rankings alone. Its enterprise scores are compelling. Its weaker results on ultra-long engineering challenges and terminal work are equally relevant.

Click to expand

What Google actually announced

Argon is initially rolling out to trusted cyber defenders through Google's Fairwind Program. Broader availability is planned, beginning with paid API customers and Google AI Ultra subscribers, but the announcement supplies no public launch date. A model announcement, an evaluator's access and general developer availability are three different things.^[1]

The most striking specification is an output token limit of one million, increased from 64,000. Google explicitly describes output capacity, including room for deep reasoning in a single trajectory. That is not interchangeable with a one-million-token input context window, nor does it promise that every task needs an enormous response.^[1]

Google also reports substantial internal engineering work: agents identifying memory optimizations and assisting C/C++ to Rust migrations. Those examples suggest a model intended for iterative, measurable work rather than just code completion. They remain Google's reported deployments, not independently reproduced productivity studies. The announcement itself says critical rewrites undergo automated and manual auditing, emulation testing and review before production.^[1]

Explore the published benchmark results

The following figures are from Google's comparison table, not Intueo measurements. They are percentages as displayed, but their underlying scoring rules differ. There is no defensible overall average across these rows.^[2]

Google's methodology says Argon generally uses the highest thinking settings and pass@1 unless otherwise specified. Competitor results largely come from provider reports or benchmark leaderboards, rather than a single controlled rerun. DeepSWE and Terminal-Bench Argon scores are self-computed; knowledge-work figures are attributed to outside evaluators.^[3]

That makes this a useful capability map, not a universal ranking. In particular, FrontierSWE's mean@5 aggregate reward is not the same measurement as a single-attempt pass rate. The selector should help readers compare models within one benchmark, not compare percentages between unrelated tests.

The coding verdict changes with the job

DeepSWE v1.1 is Argon's strongest headline: 77.9%, ahead of Opus 5.5 at 74.2% and Astra at 74.1%. Google used a mini-swe-agent harness for Argon. Datacurve's benchmark has 113 tasks; v1.1 retains the original engineering tasks but grades committed patches in clean, isolated environments and fixes dependency drift and flaky tests.^[3][4]

The distinction matters. This is DeepSWE v1.1, not SWE-bench Verified, SWE-bench Pro or Google's earlier Gemini evaluations. Its 113-task corpus provides the task denominator, but the sources do not establish the number of completed Argon trials behind Google's 77.9%. Turning that percentage into a literal number of solved tasks would invent precision.

For repository maintenance with clear behavioral tests, Argon therefore deserves attention. A sensible pilot would ask it to implement a feature, run regression tests and leave a reviewable patch. The benchmark supports that evaluation priority; it does not prove that unfamiliar proprietary repositories will achieve the same success rate.

FrontierSWE v2 tells a different story. The independent leaderboard directly reports Argon at 55.0%, below Astra's 65.5% and Opus 5.5's 62.3%. Its methodology covers 34 tasks, five trials per task and a 20-hour budget, with maximum reasoning effort and Proximus as the default harness. Scores aggregate task rewards between zero and one, rather than counting only fully solved issues.^[5][6]

These tasks include scientific computing, system implementation and performance optimization. If the work resembles building a simulator, porting a complex subsystem or designing experiments over many hours, Astra and Opus remain important comparison candidates. Google's own Terminal-Bench 4.0 table reinforces the point: Argon's 57.4% trails Opus at 66.4%. Strong repository engineering does not automatically imply the strongest terminal agent.^[2]

Enterprise work is the clearest reason to pay attention

Argon leads the displayed Vals Index by 1.9 percentage points over Opus, and AutomationBench by 8.8 points. Google describes the Vals Index as weighting finance, coding, legal and tax work by their contribution to U.S. GDP. Its methodology identifies AutomationBench's private set and attributes the score to Zapier's public leaderboard.^[1][2][3]

That combination is more interesting than an abstract intelligence claim. It suggests a candidate for workflows that connect research, documents and actions: preparing a financial analysis from multiple sources, drafting a legal research memo for professional review, or assembling a business deliverable across applications.

However, a 51.3% AutomationBench score is not evidence of dependable unattended business operations. Even a leading result leaves substantial room for failure. For a pilot, separate drafting from execution: check source citations and calculations first, then require approval before modifying records, sending messages or committing financial decisions.

The table also reports 65.4% on Vals Finance Agent v2 and 19.6% on Harvey's Legal Agent Benchmark. These are different scales and task structures, not percentages of a profession that AI can replace. Their practical value is in identifying where to run a controlled comparison against a current workflow, with domain experts judging completeness and consequential errors.^[2]

Long inputs and video offer a second advantage

On GraphWalks' 256k to one-million-token subset, Argon scores 84.2% BFS F1 versus Astra's 71.8%. Google's methodology identifies 200 problems in this subset and says all models' results were self-computed. This tests structured long-context reasoning, not just whether a model can accept a large prompt.^[2][3]

For teams working with large document collections or dependency-heavy technical material, that is a promising signal. But graph traversal is not contract review or incident analysis. A practical test should include questions whose answers require connecting distant pieces of evidence and should score whether the model cites the right passages.

LVBench adds a visual angle: Argon records 91.7%, ahead of Astra at 87.5%. Google ran these evaluations without tools, but used different sampling limits: Gemini at one frame per second, Astra at 800 frames, Fable at 300 and Opus at 600 because of API limitations. The result describes the evaluated systems and input allowances, not a perfectly identical video feed.^[2][3]

That makes long-video review worth exploring, especially when identifying an event requires earlier context. It does not establish real-time performance or justify skipping human verification of consequential observations.

Independent evidence and the real cost of reasoning

Artificial Analysis provides a separate record for Argon at high reasoning: 53 on Intelligence Index v4.3.2. That index combines ten evaluations. It also reports 110 million output tokens across its evaluation and a weighted cost of $1.99 per Intelligence Index task. These are independent measurements, not Google's Vals score under another name.^[7]

The record supports the view that Argon is a serious frontier candidate, while showing why token price alone is insufficient. A model can be competitively priced per token and still consume many tokens on a difficult task. Artificial Analysis lists no output-speed measurement, so this evidence does not establish low latency.

Google's introductory rates are $2 per million input tokens and $10 per million output tokens, with cached input priced 95% below the input rate. The announcement's pricing footnote specifies $4 input and $20 output after the introductory period, without naming its expiration date. At the introductory rate, one million billed output tokens would cost $10 before input and other applicable charges. Output headroom is capacity, not a free reasoning budget.^[1]

FrontierSWE reports $129.36 and 10.6 hours per Argon trial, explicitly warning that its cache hit rate was unusually low in a non-production setting. Those figures illustrate evaluation expense, not a production quote. Artificial Analysis also lists an API provider, but that listing should not override Google's explicit gated rollout.^[5][7]

Our verdict: shortlist selectively, deploy cautiously

Argon's strongest case is enterprise research and document-heavy reasoning, followed by testable repository engineering. For ultra-long technical builds and terminal-heavy tasks, the published evidence favors keeping Astra and Opus in the comparison. For low-latency interactions, the sources checked here do not establish a speed advantage.

Defensive security is another promising but bounded use. Argon ties Astra at 68% on CWE-bench v1; Google's methodology says ties are broken using pass@4, so equal pass@1 scores do not settle the final ordering. Authorized vulnerability remediation still needs isolation, patch review and regression testing, regardless of model rank.^[2][3]

The right next step, when access permits, is a small evaluation using representative tasks and the same tools, permissions and budget for each candidate. Measure successful outcomes, reviewer effort, elapsed time and total cost. Argon has earned a place on that shortlist. It has not earned a blanket claim of coding supremacy or production readiness from benchmarks alone.

Sources checked October 6, 2026. This review analyzes published evaluations; Intueo did not run Argon benchmarks.