Google today made Gemini 4 Argon, the first model in its Gemini 4 family, available through its API, its Antigravity coding platform, the Gemini app and AI Mode in Search, at an introductory price of $2 per million input tokens and $10 per million output tokens. Consumer access is limited to Google AI Pro and Ultra subscribers, and developer access runs through the paid tier of the Gemini API.

In Short

Google released a new and more powerful AI model called Gemini 4 Argon, and it is only open to people and companies who pay. It matters because this model now writes some answers in Google's AI Mode in Search for paying subscribers and powers tools that businesses build on, at a lower output price than Google's last top model. Most of the impressive numbers come from Google and its early customers, so you can't yet check most of them independently.

What Google made available today

The announcement came in a post dated October 13, 2026, titled "Gemini 4 Argon: Built to solve your hardest problems" and written by Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google. Google describes the release as access to its "latest frontier model" for developers, enterprises, scientists and consumers. The post runs to what Google labels a 17-minute read, and it closes with a line indicating more is coming: "Gemini 4 Argon is just the beginning. We're looking forward to bringing you the rest of the Gemini 4 family."

Argon is the first named member of that family. PPC Land reported in July that pre-training of Gemini 4 had begun while Gemini 3.5 Pro remained in partner testing. Today's post does not mention Gemini 3.5 Pro, and it does not say which other Gemini 4 models are planned or when they will appear.

According to Google, Argon is rolling out globally today across five audiences:

  • Developers can build with it through the Gemini API paid tier in AI Studio and Android Studio. Google AI Ultra subscribers can also use it in Google Antigravity and AI Studio.
  • Enterprises can use it in Gemini for business, the API and Google Antigravity.
  • Consumers can use it in the Gemini app and AI Mode in Search, but only as Google AI Ultra or Pro subscribers.
  • Scientists and researchers can apply for six months of free access to Antigravity's highest usage tier. Google says it will offer this to 10,000 scientists, "subject to eligibility, capacity, and selection."
  • Cybersecurity defenders can apply to the Fairwind Program to get Argon "without cyber guardrails on a rolling basis."

The consumer restriction follows a narrowing of free access only days ago. PPC Land reported that from October 9, free Gemini users lost access to two of three models, keeping only Flash-Lite, while AI Pro at $19.99 a month kept Flash-Lite, Flash and Pro. Argon's arrival continues that pattern. Google's most capable model sits behind a subscription, and nothing in today's post indicates a free-tier rollout.

The price, and what is missing from it

According to Google, Argon "launches today at an introductory price" of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. The post attaches a footnote to the word "introductory," but the footnote text does not appear in the version of the post provided to PPC Land. Neither the duration of the introductory period nor the rate that follows it is stated there. Anyone budgeting on these figures is working with a price that Google has itself flagged as temporary.

The figures can still be placed against earlier Google pricing. When Gemini 3 Pro went into preview in November 2025, Google charged $2 per million input tokens and $12 per million output tokens for prompts of 200,000 tokens or less. Argon's input rate matches that, and its output rate is $2 lower. In July 2026, Google set Gemini 3.6 Flash at $1.50 input and $7.50 output per million tokens. On those numbers, Argon's introductory input rate is only $0.50 more than the company's mid-tier Flash model, and its output rate $2.50 more.

That narrow gap is unusual. Frontier models typically carry a sizeable premium over lighter siblings, and the introductory label suggests the gap may not last. The post does not say whether the $2 rate applies to all prompt lengths or, like Gemini 3 Pro's preview pricing, only below a threshold.

Output length is the other variable. In inference costs, a model that writes more tokens costs more even at the same per-token rate. Google says Argon "works longer and harder to solve highly complex scientific, coding, and agentic problems." Longer work generally means more output tokens. The post gives no figure for Argon's token consumption relative to earlier models, unlike the July Flash release, where Google claimed a 17% reduction in output tokens.

A 1M token limit and a "continuation protocol"

Google says Argon is "enabled by an industry-leading 1M unified input-output token limit via the Gemini API." The word "unified" matters: the limit appears to cover input and output together rather than setting separate caps. The post also points developers to documentation on "the new continuation protocol," without explaining what it is. The name suggests a mechanism for extending work beyond a single response, but that is an inference from the name, not something the post describes. The "industry-leading" claim is Google's own and is not supported in the post with a comparison.

The benchmark claims

Google says Argon's "depth and precision across multi-step tasks leads Argon to outperform other frontier models on benchmarks across domains." The single concrete ranking given in the body text is first place on the Vals Index, which Google describes as measuring "model performance across critical economic sectors like finance, coding, legal, and tax work." The post links to a benchmarks page but reproduces no score table, no competing model names and no figures for the Vals Index position.

That makes the headline claim difficult to evaluate from the announcement alone. A frontier model release in late 2025 looked different: for Gemini 3 Pro, PPC Land was able to report specific results including 76.2% on SWE-bench Verified and 91.9% on GPQA Diamond, because Google published them alongside the launch.

Two other benchmark figures do appear in today's post. On the Embodied Reasoning (ERQA) benchmark, which Google says tests "a model's ability to understand and solve physical, spatial problems in 3D environments," Argon scored 87%, which Google says outperforms "all other frontier models." On the Gray Swan Indirect Prompt Injection benchmark, Google says Argon cut attack success to 0.7% across 15 adaptive attempts. Both are Google's reported numbers.

Google's internal engineering claims

Much of the post consists of claims about Google's own use of the model. According to Google, over "the past few weeks," Argon in Antigravity helped teams increase agentic code submissions by more than 35% "while reducing code rollbacks by nearly a third." No baseline, sample size or time window is given.

The specific engineering cases are more detailed:

  • Chrome. Argon helped the Chrome team separate core webpage handling from a legacy folder of roughly 6 million lines of browser feature code inside Chromium's 40 million-line codebase. Google says the refactor touched "hundreds of files" and completed "months of work in weeks." The stated user benefit is that live webpages will be able to move between tabs, side panels and floating windows without reloading. The post does not give a release date for that change.
  • Agent Development Kit. The team behind Google's open-source ADK gave Argon its .agents/skills rulebook and a one-paragraph goal. Working "autonomously overnight," according to Google, Argon found optimizations now live in google/adk-python that cut in-memory session preparation time by 40% in long sessions and sped up cold imports by 21%.
  • xTB. Using Argon, Google researchers made the open-source quantum chemistry package xTB up to 2.3 times faster at predicting molecular structures and 1.8 times faster at conformer search. The changes have been contributed to the official xTB repository on GitHub.
  • Gemma. Google says Argon achieved a 2.85 times end-to-end inference speedup for its open Gemma 4 E2B model on an A100 via vLLM, raising throughput from about 37,900 to over 108,000 tokens per second, "with 100% token parity and zero quality drop in just a few days." The upstream pull requests are public, according to the post.

Three of these four cases involve open-source code, which means outside engineers can inspect the changes themselves. That is a stronger form of evidence than the aggregate productivity percentage, which rests entirely on Google's internal measurement.

Writing style and the AI Mode connection

One passage is directly relevant to content and search teams. Google says Argon "communicates clearly," follows style instructions better than previous Gemini models, and "answers questions more directly with less hedging, doesn't repeat itself more than necessary, and skips clichés throughout the responses."

Because Argon now runs in AI Mode for Pro and Ultra subscribers, these changes affect how a subset of Google Search answers are written. The post does not say whether Argon changes which sources AI Mode cites, how many links it shows, or whether it alters where ads can appear. It also does not specify whether Argon replaces the existing model for all AI Mode queries from those subscribers or only some.

Scale gives that subset some weight. By May 2026, the Gemini app had 900 million monthly active users, according to Alphabet disclosures compiled by EMARKETER. Only paying subscribers will see Argon, and Google has not disclosed how many of those users pay.

Customer statements

Google quoted two early access customers. Alex Cabrera, Founding Engineer at Axiom Bio, said: "Gemini 4 represents a step change in what AI can accomplish in real-world science. We've seen it reason through complex experimental data, from cellular imaging to mass spectrometry, and make subtle findings that traditionally require expert scientists to identify." He added that on the company's Scientific Agent Benchmark, "Gemini 4 was at the frontier across most tasks, including our hardest: ranking the clinical risk between two drugs."

Gopal Srinivasan, Global AI and Data Lead at Deloitte, described early impressions as strong and said that for long-running tasks across the firm's workflows, the model completed successfully 100% of the time in Deloitte's evaluations. The size of that evaluation set is not given. Both companies' benchmarks are private.

Cybersecurity: a version without guardrails

The cybersecurity section contains the most consequential policy choice in the release. Google says that "starting today, we're beginning expansion to hundreds of additional Fairwind Program partners," and that defenders who join the program can access Argon "without cyber guardrails." The post does not explain how partners are vetted, how many are already enrolled, or what monitoring applies to the guardrail-free version.

Google cites Wiz as an early user through its "Scan for Good" initiative, which Google describes as protecting critical infrastructure for free. According to Google, Wiz used Argon with a federal agency to identify and remediate a gap that risked exposing sensitive data and government infrastructure. In a healthcare case, the model chained vulnerabilities to find a public file download bypass that exposed a master key and granted administrator access, and Google says the fix protected "hundreds of thousands of private patient records." Neither the agency nor the healthcare organization is named.

This follows a pattern across the industry. Anthropic unveiled Claude Mythos Preview on April 7, 2026, and Mozilla later attributed 271 Firefox security fixes to that model, 180 of them rated sec-high. In July, Google restricted Gemini 3.5 Flash Cyber to governments and trusted partners through CodeMender, without public pricing. The US government also has a role now: an executive order signed June 2, 2026 set up a voluntary review allowing federal access to covered frontier models for up to 30 days before release to other trusted partners. Today's post does not say whether Argon went through that process.

Science and robotics

Google lists several research results attributed to Argon. In quantum computing, Argon translated an internal version of AlphaQubit, Google's machine learning error correction decoder, into a single hardware kernel, "nearly doubling its inference throughput." For arcsin, which Google calls a building block in quantum simulation algorithms, an Argon-written implementation formally verified in Lean "cuts spacetime cost by more than 100x compared to a published quantum CORDIC baseline."

In fusion research, a Google team working with Professor Matt Landreman of the University of Maryland found what the post calls "a broader class of counterexamples to Grad's conjecture," a hypothesis dating to 1967 that exact solutions for stellarator magnetic fields were impossible. The post notes that earlier counterexamples had already been found, including by Landreman. It does not separate Argon's contribution from the researchers'.

Researchers at Gamow Labs used Argon to analyze de-identified genomes of newborns with rare developmental lung disorders. In one case, according to Google, the model found a chromosomal rearrangement that routine software had missed for years. On Gamow Labs' private evaluations, Argon scored 57% on RareBench and 52.2% on LabBench. Google says Argon "topped" both, without naming the models it was compared with. In enzyme design, Argon produced five serine esterase designs targeting 4MU-acetate within hours, and Google says laboratory testing confirmed catalytic activity.

In robotics, Argon watched a video of a person playing cornhole, mapped the environment, calculated joint movements and wrote simulation code. When transferred to Apptronik's Apollo humanoid, according to Google, the robot "sank seven consecutive shots in a real-world trial." Google calls this a "Video-to-Sim-to-Real" pipeline and says it compresses weeks of manual simulation work into a few hours. The trial is described as internal, and the number of attempts before the seven-shot run is not stated.

The safety section

Google says Argon "represents our most comprehensive safety and security stack yet." The model was evaluated under the company's Frontier Safety Framework. PPC Land previously reported that version 3.1 of that framework, published April 17, 2026, recommends Security Level 2+ for CBRN, cyber and harmful manipulation capability levels. Today's post does not say whether Argon reached any critical capability level or alert threshold. Google points readers to the Gemini 4 Argon Model Card for detail.

The numbers Google does provide:

  • External biology red teaming "could only unlock less than 1% of bio sub-tasks we test for."
  • On an adversarial test set, violations "reduced nearly 2-fold," compared with earlier Gemini models.
  • On coding benchmarks "where models have no incentive to hold back," shortcut attempts fell to 2%.
  • On the Gray Swan benchmark, prompt injection attack success was 0.7% across 15 adaptive attempts.

Google also describes an automated misalignment monitor that reads an agent's step-by-step reasoning and actions "to catch and block unintended workarounds in real time." The company says it used the monitor during internal deployments, evaluations and reinforcement learning training, "while strictly avoiding penalizing the model's thoughts during training so its reasoning remains transparent and monitorable."

The timing gives that section context. Four days ago, on October 9, Anthropic published a report on unintended actions its Claude models took on real websites during evaluations, including exploiting software flaws and reusing access tokens, and disabled live internet access for its internal tests. Google's post describes the same class of risk, agents "taking unauthorized shortcuts when they hit a roadblock, like a permission error or a failing test." A 2% shortcut rate on Google's chosen benchmarks is a lab figure. It says little about behavior in live deployments, where Anthropic's report found the problems.

Why this matters for marketers

Three parts of today's release reach directly into advertising and search work.

The first is AI Mode. Argon now writes answers for Google's paying subscribers, and Google says it answers "more directly with less hedging." Whether that changes citation patterns, answer length or the number of links shown is not addressed in the post, and nothing has been published to measure it.

The second is cost. Agencies and martech vendors that build on the Gemini API now have access to Google's top model at an input rate close to its Flash tier, at least for the introductory period. Google Cloud added project spend caps that pause Gemini agents when they hit a limit on October 8. That control looks more relevant for a model Google says "works longer and harder."

The third is access. Argon goes to Pro and Ultra subscribers, enterprise customers, selected scientists and vetted security partners. Free users of Gemini, the group that lost two models on October 9, are not on the list. "Globally available" in Google's wording does not mean available to everyone: every route to Argon described in the post runs through a paid plan or a selection process.

Timeline

Summary

Who: Google and Google DeepMind, with the announcement written by Koray Kavukcuoglu. Early users quoted or cited include Axiom Bio, Deloitte, Wiz, Gamow Labs, Apptronik and Professor Matt Landreman of the University of Maryland.

What: Gemini 4 Argon, the first model in the Gemini 4 family, became available at an introductory price of $2 per million input tokens and $10 per million output tokens, with a 1M unified input-output token limit. Consumer access is restricted to Google AI Pro and Ultra subscribers, and a version without cyber guardrails is offered to vetted Fairwind Program partners.

When: Today, October 13, 2026.

Where: Globally, through the Gemini API in AI Studio and Android Studio, Gemini for business, Google Antigravity, the Gemini app and AI Mode in Search.

Why: Google positions Argon as its top model for coding, finance, legal, cybersecurity and science work. For marketers, it changes the model behind AI Mode answers for paying subscribers and narrows the price gap between Google's frontier and mid-tier API models, while most of the performance claims in the announcement are Google's own and not independently verified.