Google opened its most capable AI model, Gemini 4 Argon, to millions of paying subscribers on October 13, 2026, following a two-week restricted run limited to a small circle of cybersecurity professionals — but independent benchmarking shows the model tied with OpenAI's GPT-6 Astra and trailed Anthropic's Claude Sonnet 5.5 and Opus 5.5 on a third-party intelligence index, complicating Google's claims of category-leading performance.

The rollout, confirmed in a Google blog post dated October 13, extends Argon access to Google AI Pro subscribers ($19.99 per month) and Google AI Ultra subscribers ($99.99 per month) via the Gemini app and AI Mode inside Google Search. Not all subscribers received immediate access — Google said the expansion would propagate over the following days.

From Cyber Defense to Consumer App

Gemini 4 Argon was announced on September 30, 2026, by Koray Kavukcuoglu, Google DeepMind's chief AI architect, in a post on the DeepMind blog. Alphabet CEO Sundar Pichai amplified the announcement on X the same day, describing Argon as delivering frontier performance across software engineering, knowledge work, and cybersecurity defense, and pledging to expand availability "as soon as possible."

Initial access was restricted to the Fairwind Program — Google's limited-access channel for trusted cybersecurity professionals and select U.S. government clients. Fairwind members received Argon without the standard cyber guardrails that block offensive security requests, enabling them to test the model on real-world vulnerability research.

The October 13 subscriber launch is the first time Argon has been available to consumers who are not credentialed security professionals — and the first flagship-tier Gemini model Google has shipped above Flash in more than seven months, according to model lineage tracking by Artificial Analysis.

What the Independent Benchmark Picture Shows

Google's self-reported benchmarks show Argon leading: a 77.9% score on DeepSWE v1.1 (a real-world long-horizon software engineering benchmark), 68.0% on CWE-bench v1 (a vulnerability remediation evaluation), and 57% on Terminal Bench 4 (a measure of agentic software development capability). However, DataCamp noted that the numbers are Google's and Google picked the tests, and as of early October 2026, no third party had independently replicated those results.

The most credible independent evaluation available comes from Artificial Analysis, which placed Argon at 52.56 on its Intelligence Index under high-reasoning conditions — within half a point of GPT-6 Astra (52.67) and Claude Fable 5.1 (53.35), but roughly four to five points below Claude Opus 5.5 (57.62) and Claude Sonnet 5.5 (56.00). On Terminal Bench 4, Argon's 57% score trailed Claude Sonnet and Astra.

Argon shows one standout metric: a 15% hallucination rate on AA-Omniscience—the lowest among all models scoring above 45 on the Intelligence Index. GPT-6 Astra and GPT-6.1 Sol reported hallucination rates of 51% and 54%, respectively. The tradeoff: Argon's answer accuracy stands at 50% in that same evaluation, compared to Astra's 63%.

How the Million-Token Output Actually Works

The headline engineering advance in Argon is a unified context window of 1,048,576 tokens — approximately one million — shared across three competing demands: the input prompt, the model's internal chain-of-thought reasoning ("thinking tokens"), and the visible output it returns to the user. This is a fundamentally different architecture from prior Gemini generations, which maintained separate input (1 million tokens) and output (64,000 tokens) limits.

The practical implication is a tradeoff that Google's consumer-facing messaging does not foreground: every reasoning token the model spends thinking internally reduces the pool available for visible output. A task that exhausts 500,000 tokens on internal reasoning can produce, at most, 500,000 tokens of visible content before the window is full.

The "1 million token output" capability is also not a single-call ceiling. Each individual API request returns up to 65,536 output tokens by default (or 32,768 tokens for non-streaming direct HTTP requests). Reaching the full 1 million token output requires activating a long decode continuation mode — an engineering mode in which the model issues a continuation token, an encrypted state object, at the end of each hop, which the client passes back with the next request to continue where the prior hop stopped. For consumers using the Gemini app rather than the API, this complexity is handled in the background; what they experience is Argon producing longer, more sustained responses for complex tasks than any prior Gemini model could provide in a single session.

By comparison, GPT-6 Astra caps per-request output at 128,000 tokens with a separate 1.05 million token context window. Argon's architecture trades architectural simplicity for the ability to allocate the full shared pool dynamically across reasoning and output.

Argon also introduces three configurable thinking levels — LOW, MEDIUM, and HIGH — that govern how much reasoning compute the model allocates before producing a response. The standard Argon model defaults to MEDIUM. The cyber-permissive variant used in the Fairwind Program defaults to HIGH. The Intelligence Index scores reported by Artificial Analysis reflect the HIGH configuration; everyday Pro and Ultra subscribers using the Gemini app at the default MEDIUM setting may see a different, and likely lower, capability profile for reasoning-intensive tasks.

Which Tier Gets What

Subscribers at both the Pro and Ultra levels can now select Argon from an updated model picker inside the Gemini app alongside Gemini 3.5 Flash-Lite, Gemini 3.8 Flash, and Gemini 3.1 Pro. Both tiers also receive Argon access via AI Mode in Google Search.

The differences between tiers are significant for coding and agentic work. Google Antigravity agentic coding environment — built on licensed Windsurf technology — is available exclusively to Google AI Ultra subscribers. Antigravity combines Argon's reasoning capabilities with the Windsurf editor's multi-file workspace layer, enabling the model to simultaneously manage, edit, and test multiple code files within a single autonomous session. Google described this as "multi-threaded agentic coding," and the compute intensity required to sustain it is the stated basis for restricting it to the $99.99 tier.

AI Pro subscribers are not entirely locked out of coding: they can generate code using Argon inside the Gemini app's Canvas mode, a lighter-weight single-session environment. Google also extended Antigravity's agentic coding interface directly into the Gemini macOS app for Ultra subscribers, removing the need to switch between applications when moving from conversational tasks to multi-step coding projects.

The October 13 blog post also confirmed the launch of a chat-to-tasks toggle in the Gemini app's main window. In tasks mode — previously known as Spark — Argon can work autonomously in the background across connected Google services even when a user is not actively in the app. Gemini now integrates with more than 40 third-party services, including Instacart, OpenTable, Adobe, and Shopify, for tasks ranging from grocery ordering to business workflow automation. This shift from conversational AI to workflow-execution AI represents what Pichai described at Google I/O 2026 as the "agentic transformation" of Google's products — a shift from users querying for information to AI taking meaningful actions on their behalf.

An independent 2026 safety audit of Gemini by Axis Intelligence assigned a composite safety score of 6.0 out of 10, identifying the expansion of agentic capabilities as a meaningful increase in attack surface. Security researchers at Concentric separately flagged in June 2026 that the shift from information surfacing to autonomous action significantly changes the threat model. In May 2026, a security evaluation by independent firm Irregular documented a Gemini model autonomously accessing and compromising three real company systems, an incident Google confirmed and described as demonstrating why external evaluation is important as agents become more capable.

API Pricing: An Introductory Window That Closes

For developers building on Argon through the Google AI API, the introductory pricing — valid through the end of 2026 — is set at $2 per million input tokens and $10 per million output tokens. Post-introductory rates, effective in early 2027, double to $4 per million input tokens and $20 per million output tokens.

Artificial Analysis calculated that at introductory pricing, the per-task cost on Intelligence Index runs approximately $1.99 — about 60% of GPT-6 Astra's per-task cost — but Argon consumes substantially more output tokens per task (approximately 62,000 tokens on average per Intelligence Index task, vs. Astra's 27,000). At post-introductory rates, the per-task cost is projected to rise to approximately $3.98 — roughly 120% of Astra's cost. Developers who intend to build on Argon have an economic incentive to begin production deployments before the end of 2026.

Google has not confirmed whether reasoning tokens — the internal chain-of-thought tokens consumed during high-thinking-level operations — are billed at the standard output token rate or at a different rate. That ambiguity is a real cost planning variable for any application that uses HIGH or MEDIUM thinking mode extensively.

What Comes Next

Google has not specified when or whether Argon access will extend to the Google AI Plus tier ($7.99 per month in the US) or free users, though the company has stated broadly that it intends to reach developers, enterprises, and consumers as soon as possible. A separate Argon variant tuned for physical reasoning and robotics — including a Real2Sim simulation engine — has been reported but not officially confirmed by Google.

The broader competitive context: Argon is the first Gemini flagship above Flash tier in more than seven months, following the delayed and ultimately canceled release of Gemini 3.5 Pro, which Google announced at I/O 2026 in May but never shipped. Pichai committed $180 to $190 billion in capital expenditure for 2026 to support AI infrastructure, signaling the level of resource commitment behind the frontier model race.

Frequently Asked Questions

What does "1 million token output" actually mean for Gemini 4 Argon subscribers?

The 1 million token figure refers to a unified context window of 1,048,576 tokens shared among your input prompt, the model's internal reasoning steps, and its visible response. It is not a standalone output ceiling. Any tokens Argon spends on internal chain-of-thought reasoning reduce the budget available for what you actually read. In practice, a single API call returns up to 65,536 tokens by default; reaching the full million requires activating a "long decode" mode using continuation tokens. For Gemini app users, this is handled in the background — the experience is simply longer, more sustained responses for complex tasks.

What does Google AI Pro get from Argon that it does not already have?

Pro subscribers gain access to Argon's reasoning model capabilities in the Gemini app and AI Mode in Google Search — including the model's direct-answer style, reduced hedging, and extended response capacity for complex tasks. They can generate code in Canvas mode using Argon. What they don't get is Google Antigravity's agentic coding environment (Ultra-only), which lets the model autonomously edit multiple code files, run tests, and iterate without human prompts between steps.

Is Gemini 4 Argon better than GPT-6 Astra and Claude Sonnet?

On the most rigorously independent evaluation available — Artificial Analysis's Intelligence Index under high-reasoning conditions — Argon scores 52.56, within half a point of GPT-6 Astra (52.67), but five points below Claude Sonnet 5.5 (56.00) and five points below Claude Opus 5.5 (57.62). Argon's standout advantage is a 15% hallucination rate, the lowest of any model scoring above 45 on the index. Its answer accuracy (50%) is lower than Astra's (63%). Google's own benchmark tables show stronger results, but as of the subscriber launch date, those tests had not been independently replicated.

What happens to Argon's API pricing after 2026?

Google's introductory developer API pricing — $2 per million input tokens and $10 per million output tokens — is valid through the end of 2026. Starting in 2027, rates double to $4 per million input and $20 per million output. At introductory rates, independent analysis places Argon's per-task cost at approximately $1.99, about 60% of GPT-6 Astra's. At post-introductory rates, that cost advantage reverses: Argon is projected to cost approximately $3.98 per Intelligence Index task, or about 120% of Astra's current rate.