Google's Gemini 4 Argon Has Impressive Benchmarks, and Some Important Caveats

Google's new flagship model went from cyber-defender-only access to general availability in 13 days. The benchmark story behind it is messier than Google's own chart suggests.

On September 30, 2026, Google handed its newest frontier model to a small group of cybersecurity partners and almost no one else. Thirteen days later, on October 13, that same model, Gemini 4 Argon, became generally available: paid API access for developers, plus the Gemini app for Google AI Pro and Ultra subscribers (Google Cloud Blog; Gizmodo). That phased, security-first rollout, cyber defenders first and the general public thirteen days later, is a departure from how Google has typically shipped past Gemini updates.

What makes Argon worth paying attention to isn't just the rollout, though. It's the gap between what Google's own benchmark chart claims and what independent testers have measured since. Google says Argon leads the frontier on a majority of its published benchmarks. Independent evaluations from firms like Artificial Analysis and Vals AI back that up on specific business and legal-reasoning tasks, but show a closer, more mixed picture on general reasoning and terminal-based coding, closer to "tied with GPT-6 Astra" than "ahead of the field."

What Gemini 4 Argon actually is

Google's model card describes Argon as a new frontier model rather than a fine-tune or modification of an earlier Gemini release. Its training process includes pre-training across diverse data, followed by post-training techniques, including reinforcement learning aimed at multi-step reasoning, problem-solving, and theorem-proving (Google DeepMind).

The headline technical specs, per Google's developer documentation:

• Context window: 1,048,576 tokens (roughly 1 million), shared across input, internal "thinking," and output

• Max output: 65,536 tokens by default for a standard synchronous response; up to the remainder of the shared 1,048,576-token budget using a "long decode" mode, which generates in 64K-token hops that are chained together via continuation requests rather than returned as one instant response

• Inputs supported: text, image, video, audio, and PDF; output is text only

• Reasoning modes: low, medium, and high "thinking" settings

• Other capabilities: code execution, computer use (in preview), file search, function calling, structured outputs, search grounding, and grounding with Google Maps (Google AI for Developers; Google AI for Developers, long decode docs)

It's worth being precise about that last point, because Google's own marketing shorthand ("1M output tokens") oversimplifies it. Argon doesn't have a separate million-token output allowance on top of its context window; input, reasoning, and output all draw from the same 1,048,576-token pool, so a long prompt leaves less room for the response. And a response anywhere near that size isn't generated in a single pass: long decode works by producing output in 64K-token hops, each billed separately, stitched together through a continuation mechanism rather than arriving all at once (Google AI for Developers, long decode docs). The practical upshot is still significant: it's a large jump from the roughly 64K-token output ceiling on Gemini 3.1 Pro, and it's aimed at long-horizon jobs like large-scale codebase migrations and lengthy legal documents that previously had to be chunked across many separate requests. It just isn't the one-shot, million-token reply the headline number implies.

Argon is also notably resistant, by Google's account, to indirect prompt injection, the attack pattern where malicious instructions are hidden inside a document, webpage, or tool output that the model processes, hijacking its behavior. Google's enterprise announcement calls Argon "our most resilient model yet" against that class of attack (Google Cloud Blog). That's Google's own characterization; it hasn't been independently verified by a third-party security audit as of publication.

Why Google shipped it to cybersecurity defenders first

The rollout strategy is the part of this story that's genuinely unusual. Rather than a simultaneous public launch, Google released Argon first, and specifically for this vetted group, without the usual cyber safety guardrails, to a limited set of "trusted cyber defenders" through a program it calls Fairwind, alongside its own internal security teams (Google Blog). That guardrail-free access applied only to the Fairwind cohort; it does not describe the general-availability version of Argon that launched October 13.

CNBC's reporting on the launch quotes a Google executive, Doshi, explaining the reasoning in blunt terms: a model at this level of frontier capability is "meaningfully important for defenders." On cybersecurity-specific benchmarks, Argon reportedly ties with OpenAI's GPT-6 Astra and xAI's Grok 4.7. Separately, on the independent Vals Index, a different kind of evaluation covering business and professional tasks more broadly, Argon ranks ahead of GPT-6 Astra and recent Anthropic models (CNBC).

The subtext is that frontier-level coding and reasoning models are now capable enough to meaningfully help attackers too, and Google wanted defensive teams equipped before opening the model to the general public. Google says it is working with the U.S. government through a voluntary pre-release review process, and the model's own system card acknowledges the company "cannot rule out" that Argon reaches critical capability thresholds in both biological and cybersecurity risk categories. That kind of disclosure is now fairly standard across frontier labs' safety reports, but it's still notable to see stated plainly (Google DeepMind).

Google also says it's using Argon internally to optimize memory usage inside its own data centers, claiming the model has freed up "hundreds of terabytes" of memory without new hardware purchases. That's a specific, checkable-sounding claim, but it comes from Google itself; no independent party has verified the figure (CNBC).

The rollout, mapped out

• September 30, 2026: Announced; limited to trusted cybersecurity partners via the Fairwind Program, plus internal Google teams (Google Blog)

• October 13, 2026: General availability across the channels Google announced that day: the paid Gemini API tier and Google AI Studio for developers, Google Antigravity and Android Studio, Google Cloud's Gemini for business, and the consumer Gemini app for Google AI Pro and Ultra subscribers (Google Cloud Blog; Google Blog, Gemini app; Google Blog, developers; Gizmodo)

"General availability" means the paid tiers and subscriptions listed above are open to anyone who qualifies for them; it isn't unrestricted free access, and eligibility still varies by product. Individual consumer-facing features are rolling out gradually from there: Google AI Pro access to the Gemini app, for instance, is expanding country by country rather than appearing everywhere simultaneously, and some robotics-specific preview features remain limited to Google AI Ultra subscribers (Google Developers Blog; Layer3Labs).

That's 13 days from a cyber-defenders-only preview to general availability, longer and more deliberate than Google's past practice of shipping new models to everyone simultaneously, though faster than some early coverage anticipated. In the gap between the two dates, several outlets noted the awkwardness of a company publishing benchmark claims almost nobody could test firsthand; one Dev Community piece captured that mood with the headline "Gemini 4 Argon tops the benchmarks. You can't use it" (DEV Community). As of October 13, that headline is no longer accurate.

Pricing: aggressive, for now

Google's own pricing documentation, published alongside the October 13 release, lists an introductory API rate of $2 per million input tokens and $10 per million output tokens through November 30, 2026, with cached input tokens discounted 95%. Standard pricing from December 1, 2026 is listed at $4 per million input tokens and $20 per million output tokens (Google Cloud Blog).

That introductory rate undercuts both GPT-6 Astra and Claude Opus 5.5 on paper. Separately, Vals AI, the independent evaluation firm behind the Vals Index, reports a cost of roughly $15.68 to run its full evaluation suite against Argon at standard ($4/$20) pricing, compared with about $32.14 for Claude Opus 5.5 (Vals AI; cited via DEV Community). That's a benchmark-specific cost figure, not a forecast of what an ordinary production workload will cost; real-world bills depend heavily on token usage, caching, and workload mix. Still, it points in the same direction as the headline pricing: Argon's combination of competitive scores and lower per-token cost looks like a real part of Google's pitch, not just a launch-week promotion.

Businesses modeling costs around Argon should plan for the step up to standard pricing on December 1. Google has already published that increase; it isn't speculative. Per-token rates will roughly double, though actual bills will vary by how each team uses the model.

The benchmark picture: strong, but not the clean sweep Google's chart implies

This is where the story gets genuinely interesting, because the gap between Google's self-reported numbers and independent testing is wider than usual.

Google's own table, checked against the published model card, shows Argon ahead of GPT-6 Astra and Claude Opus 5.5 on several fronts: 77.9% on DeepSWE v1.1 (versus 74.1% and 74.2%, respectively), 51.3% on AutomationBench (versus 41.4% and 42.5%), a strong showing on long-context retrieval (GraphWalks), and a reported edge on Harvey's legal-agent benchmark (Google DeepMind). It trails on others, by Google's own numbers: on FrontierSWE v2, Argon's 55.0% is last among the four models compared, behind GPT-6 Astra's 65.5% and Claude Opus 5.5's 62.3%; on Terminal-Bench 4.0, Argon's 57.4% trails Claude Opus 5.5's 66.4%. Across the roughly 19 benchmark rows Google published, independent tallies put Argon's record at about 13 wins, one tie, and five losses against the three comparison models, with the losses concentrated in terminal-driven and shell-based coding tasks (Apidog; DataCamp).

Independent testing tells a flatter story on general reasoning. Artificial Analysis's Intelligence Index, one of the more widely cited third-party aggregates, puts Argon at 53 points at its highest ("high") reasoning setting, checked as of mid-October, essentially tied with GPT-6 Astra and roughly five points behind Claude Opus 5.5's 58 (R&D World; Emergent). As with every figure in this section, reasoning settings and test dates affect the score, so treat it as a snapshot rather than a fixed ranking.

Where Argon does appear to hold a real, independently measured edge:

• Vals Index. In Vals AI's evaluation shortly after launch, Argon ranked first among the models tested at that time, at 68.9%, reportedly the first Gemini model to top that index (Emergent; MindStudio). Vals AI's leaderboard has since expanded as more models are added, so treat the ranking as a snapshot from its initial post-launch evaluation rather than a fixed, permanent standing.

• AutomationBench-AA. This is a separate evaluation from Google's own AutomationBench score above, run independently by Artificial Analysis rather than Google, despite the similar name. On that independent test, Argon ranks first at 78%, ahead of Claude Sonnet 5.5 (Gemini4.link).

• Hallucination rate. On Artificial Analysis's AA-Omniscience benchmark, Argon posts a 15% hallucination rate, defined as the share of wrong answers among the questions it actually attempted, excluding ones it declined to answer. That's the lowest rate recorded among models scoring 45-plus on the Intelligence Index as of this evaluation. The nuance matters: it isn't the same as overall accuracy. Argon answers correctly on about 50% of questions on this test, versus 63% for GPT-6 Astra, but it declines to answer far more often instead of guessing, which is why its share of confidently wrong answers stays low even though its raw hit rate trails Astra's (Artificial Analysis; Gemini4.link).

Where independent results are less flattering than Google's chart alone might suggest:

• Terminal and shell-based coding agent work. Independent analysis generally confirms Argon trails Claude Opus 5.5 on Terminal-Bench-style tasks and GPT-6 Astra on broader software-engineering suites. One review, based on published benchmark data and the author's own API testing, put it plainly: for agents that operate through a command-line shell, Argon isn't the first choice (eesel AI).

• Real-world usage versus benchmarks. A Bloomberg report cited in multiple early writeups suggested Argon performs well on standardized tests but underperforms when employees use it for day-to-day work, the kind of gap industry watchers have started calling "benchmaxing," where models are implicitly or explicitly tuned toward the specific tests labs publish (MindStudio, citing Bloomberg).

Google, for its part, disputes the characterization that its lead is overstated and maintains that Argon leads the frontier on a majority of the benchmarks it tracks (R&D World).

A caveat that applies to every number in this article, including the table below: benchmark results aren't always apples-to-apples. Different testers use different reasoning settings, different tool access, and different scoring methodology, and figures were checked at various points between late September and mid-October as Google and third parties kept testing. Treat these numbers as directional evidence of where Argon is strong and weak, not as a precise, reproducible leaderboard.

Gemini 4 Argon vs. the competition, at a glance

Rows are labeled by source, Google's own published table versus independently run evaluations, because the two use different methodology and reasoning settings and aren't always directly comparable.

Gemini 4 Argon GPT-6 Astra Claude Opus 5.5 Announced Sept 30, 2026 Sept 3, 2026 Sept 22, 2026 Artificial Analysis Intelligence Index (independent, high/max reasoning) 53 53 58 Vals Index (independent, initial post-launch ranking) 68.9% (1st) 63.1% 67.0% DeepSWE v1.1 (Google's table) 77.9% 74.1% 74.2% FrontierSWE v2 (Google's table) 55.0% 65.5% 62.3% Terminal-Bench 4.0 (Google's table) 57.4% 58.2% 66.4% Shared context/output budget 1,048,576 tokens, shared across input and output not published not published Introductory API price (per million tokens) $2 in / $10 out not published here not published here

Figures compiled from Google's published model card, Artificial Analysis, and Vals AI; see sources linked throughout this article. "Not published here" indicates a figure this article did not independently confirm for that model, not that no such figure exists. (Google DeepMind; NeuralTrust; Emergent)

Who should actually use it

Published benchmark results suggest long-document and knowledge work is where Argon's case is strongest: contracts, financial filings, and large codebases that need to be understood holistically rather than edited line-by-line in a terminal. The shared million-token context budget, the legal and finance benchmark results, and the low rate of confidently wrong answers all point the same direction, though none of that is a substitute for testing the model on your own workload before committing to it.

Agentic coding through a terminal or shell is where the evidence, both Google's own numbers and independent testing, consistently shows Argon behind. Claude Opus 5.5 currently has the edge there, with GPT-6 Astra competitive as well.

On cost, Argon's introductory pricing beats both rivals, but Google has already published the December 1 rate increase, so it's a launch discount rather than a durable price advantage. For security teams specifically, Argon's cyber-defender-first rollout and its reported resistance to indirect prompt injection are worth tracking regardless of whether you adopt the model more broadly, since they signal where Google sees the most urgent near-term risk from frontier models.

Gemini 4 Argon is a capable new addition to the frontier tier, and Google isn't wrong that it leads a number of important benchmarks. What the independent testing adds is texture that Google's own chart doesn't offer on its own: Argon's strength looks concentrated in specific domains, business workflows, legal and financial reasoning, and long documents, rather than spread evenly across the board. Judge it on the tasks you'd actually hand it, not on a single leaderboard position.