{"schemaVersion":"1.0","type":"Article","types":["Article"],"slug":"gemini-4-argon-is-here-what-google-actually-shipped-who-gets-it-and-what-it-costs-mx6ua","url":"https://zyvop.com/gemini-4-argon-is-here-what-google-actually-shipped-who-gets-it-and-what-it-costs-mx6ua","title":"Gemini 4 Argon Is Here: What Google Actually Shipped, Who Gets It, and What It Costs","subtitle":"Google's new flagship went from a cyber-defender-only preview to a public rollout in under two weeks. Here's what's real, what's marketing, and what it costs.","tldr":"Gemini 4 Argon is now live in the Gemini API, AI Studio, Antigravity and the Gemini app. Here's who gets it, what it costs, how it benchmarks, and where the hype deserves a grain of salt.","keywords":["Gemini 4 Argon","developer tools","Google","AI models","LLM"],"entities":["Samod Alex","Gemini 4 Argon","developer tools","Google","AI models","LLM","ZyVOP"],"keyTakeaways":["Google just opened its newest flagship model, Gemini 4 Argon, to developers, companies and paying Gemini subscribers.","The way it got here is a bit unusual.","On September 30, Google announced it but released it only to a set of trusted cyber defenders."],"headings":["The short version","How the rollout played out","Who can use it today","What it costs","What Google says it's good at","The benchmark story","What the independent evaluators say","The safety side","The backdrop","What I'd do this week","Bottom line","Sources"],"outboundLinks":["https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/","https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/","https://blog.google/innovation-and-ai/technology/safety-security/gemini-4-model-safety-security/","https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-4-argon/","https://www.khaleejtimes.com/business/google-announces-gemini-4-flagship-ai-model-after-months-of-delays-2","https://9to5google.com/2026/10/13/gemini-app-4-argon/","https://artificialanalysis.ai/articles/gemini-4-argon-google-top-three-labs","https://arena.ai/leaderboard/text","https://epoch.ai/benchmarks/deepswe/review","https://kingy.ai/blog/frontier-ai-models-gemini-4-argon-gpt-6-astra-sol-claude/","https://computingforgeeks.com/gemini-4-argon-released-features-benchmarks/"],"contentText":"Google just opened its newest flagship model, Gemini 4 Argon, to developers, companies and paying Gemini subscribers. The way it got here is a bit unusual. On September 30, Google announced it but released it only to a set of trusted cyber defenders. On October 13, it widened the rollout to the Gemini API, Google Antigravity, the Gemini app and AI Mode in Search. I read Google's launch posts and safety write-up, Reuters' coverage, and the outside evaluations I could find: Artificial Analysis, Vals, Arena and Epoch AI. This is the version I'd want to read before switching anything over. Heads up: everything below comes from Google's posts, Reuters and third-party evaluators. I haven't run my own tests on Argon yet, so read the benchmark sections as reporting, not a hands-on review. The short version What it is: Google's new frontier model and the anchor of its Gemini 4 generation. It targets long, multi-step work like large-scale coding, legal and finance tasks, and cyber defense. Who can use it: paying users only. That means the paid Gemini API tier, AI Pro and Ultra subscribers in the Gemini app, and Ultra subscribers in Antigravity. Price: $2 per million input tokens and $10 per million output tokens until November 30, 2026. After that it doubles. Where it shines: independent tests back it on knowledge-work benchmarks, long context and a low hallucination rate. Where it doesn't: several hard coding and terminal tests, and one major composite index where Claude Opus 5.5 still scores higher. It's also wordy, which matters for cost. How the rollout played out Most model launches are one big switch. This one was staged. flowchart LR A[\"Sep 30, 2026<br/>Announced<br/>Fairwind cyber defenders only\"] --> B[\"Oct 13, 2026<br/>Broad paid rollout<br/>API, Antigravity, Gemini app\"] B --> C[\"Nov 30, 2026<br/>Intro pricing ends<br/>$2/$10 becomes $4/$20\"]Google says Argon can autonomously find, validate and patch critical software vulnerabilities. Its launch post says safely releasing a model like this needs a phased approach, so cyber defenders got it first. That early access produced real examples. Security company Wiz used Argon to uncover a critical bug in hospital software that earlier models had missed. Google says that helped protect hundreds of thousands of patient records. Google's own security teams also used it to find and patch bugs in internal and open-source software. Some of those fixes, for foundational vulnerabilities that affect systems globally, are going through responsible disclosure. Google also says it took part in the U.S. government's voluntary process for pre-release model access. One more thing worth knowing. Google said it would release Argon without cyber guardrails only for trusted defenders and its own internal teams. Defenders can apply to the Fairwind Program for that version, and Google says it's expanding the program to hundreds of additional partners. Who can use it today flowchart TD Q{\"Who are you?\"} --> D[\"Developer or company\"] Q --> S[\"Subscriber\"] Q --> R[\"Researcher\"] Q --> C[\"Security defender\"] D --> D1[\"Gemini API paid tier<br/>Gemini platform, Gemini for business\"] S --> S1[\"AI Pro or Ultra<br/>Gemini app + AI Mode in Search\"] S --> S2[\"AI Ultra only<br/>Antigravity\"] R --> R1[\"Apply for free top Antigravity tier<br/>6 months, 10,000 spots\"] C --> C1[\"Apply to Fairwind Program<br/>access without cyber guardrails\"]You are Where to get it The catch Developer or company Gemini API, Gemini platform, Gemini for business Paid tier only. Model ID is gemini-4-argon Everyday subscriber Gemini app and AI Mode in Search Needs Google AI Pro or Ultra Agent-first coder Google Antigravity AI Ultra for individuals. Google also lists it for enterprises Scientist Free top-tier Antigravity usage for 6 months 10,000 spots, application, subject to eligibility Cyber defender Fairwind Program Apply, access is rolling AI Studio is a bit murky. Google's launch post lists it under both the paid API tier and Ultra, while 9to5Google reports that AI Ultra is required for AI Studio and Android Studio (Canary). Check your own account before you plan around it. If you're on the free Gemini plan: Google's announcement only names Pro and Ultra subscribers for the consumer apps. It doesn't mention a free-tier rollout, so I wouldn't count on it today. What it costs Period Input (per 1M tokens) Output (per 1M tokens) Through Nov 30, 2026 (introductory) $2 $10 After Nov 30, 2026 $4 $20 Google says cached input tokens are 95% cheaper. Kingy AI's read of Google's pricing page says that's implicit caching only, with no explicit cache option. The introductory price applies to the Argon model meant for public access. So plan for the higher number. If you're budgeting a production workload, price it at $4 and $20, and treat the first seven weeks or so as a discount. Price per token is not price per task. Argon is wordy. Artificial Analysis found it averaged about 62K output tokens per task on its index, against about 27K for GPT-6 Astra. Here's how that played out in two evaluations: Evaluation Argon GPT-6 Astra GPT-6.1 Sol Artificial Analysis, per task, Argon at launch pricing ($2/$10) $1.99 $3.26 $0.72 Artificial Analysis, per task, Argon at standard pricing ($4/$20) $3.98 $3.26 $0.72 Vals, per test, Argon at standard pricing ($4/$20) $15.68 $18.46 $3.24 So it's not clearly cheaper than Astra, and it's well above Sol on both suites. The Sol and Vals figures come through Kingy AI's roundup of the evaluators' pages. Measure cost on your own workload before you decide. Context and output limits. Google advertises a 1M unified input-output token limit in the Gemini API. According to Kingy AI's read of the API docs, input, thinking and output all share that one window. A normal request caps output at 65,536 tokens, and going longer means enabling \"long decode\". Google mentions a new continuation protocol for that, so read its developer docs before you build on it. What Google says it's good at Google leans on real engineering work rather than puzzles. These are Google's own examples, so keep that in mind. Big refactors. In Chrome, Argon helped separate core webpage handling from a roughly 6 million line legacy folder inside Chromium's 40 million line codebase. Google says it turned months of work into weeks. Language migrations. Google says Argon agents are migrating C/C++ code to Rust across the company, up to 800K+ lines for the Fuchsia Zircon kernel. For libgav1, its video decoder, Argon replaced 32K lines of SIMD code and produced a memory-safe decoder that runs 2.7x faster than the earlier Rust port, with identical video output. Google adds that these rewrites are still going through auditing and review before production. Overnight optimization. The team behind Google's open-source Agent Development Kit gave Argon a rulebook and a one-paragraph goal. Working on its own overnight, it cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%. Those changes are live in the repo. Faster inference. Google reports a 2.85x end-to-end speedup for its open Gemma 4 E2B model on an A100 via vLLM. Throughput went from about 37,900 to over 108,000 tokens per second, with 100% token parity. Science. Google researchers used Argon to make the xTB quantum chemistry package up to 2.3x faster at predicting molecular structures. Argon also helped design five new serine esterase enzymes that showed real catalytic activity in lab tests. Robotics. My favorite one: in an internal test, Argon watched a video of a person playing cornhole and built the simulation code. An Apptronik Apollo humanoid then sank seven shots in a row. Google also reports 87% on the ERQA embodied reasoning benchmark. Inside Google, the company says Argon in Antigravity helped teams raise agentic code submissions by more than 35% while cutting code rollbacks by nearly a third. The benchmark story Google's numbers look strong in some places and mixed in others. Here are selected results, all vendor-reported. Benchmark What it measures Argon GPT-6 Astra Claude Opus 5.5 DeepSWE v1.1 Long-horizon software engineering 77.9% 74.1% 74.2% Vals Index Finance, coding, legal, tax 68.9% 63.1% 67.0% AutomationBench Multi-step workflow automation 51.3% 41.4% 42.5% Harvey's Legal Agent Legal research and drafting 19.6% 5.4% 3.8% GraphWalks 256K to 1M Long-context reasoning 84.2% 71.8% 66.8% CWE-bench v1 Fixing security vulnerabilities 68.0% (tie) 68.0% (tie) 67.0% FrontierSWE v2 Harder software engineering 55.0% 65.5% 62.3% Terminal-Bench 4.0 Terminal-based agent work 57.4% 58.2% 66.4% Terminal-Bench Science 0.1 Science tasks in a terminal 57.6% 68.1% 63.3% The pattern is clear. Argon leads on business-style work and very long context. It trails on several of the hardest coding and terminal tests. Reuters said the same at launch: Argon stayed behind on two of the four coding benchmarks Google included in its own release. Two caveats on the table. First, Kingy AI's read of Google's evaluation PDF says the runs aren't perfectly apples to apples. For example, Argon's Terminal-Bench Science run had a longer verifier timeout, and LVBench used different frame budgets per model. So some gaps may not come from the model alone. Second, the headline coding result has a problem. Epoch AI reviewed DeepSWE v1.1 on September 7 and rated it \"Flawed\". It found grading errors in at least 23 of 113 tasks, mostly because agent-written tests clashed with the grader. That review predates Argon and applies to every model, but a 3.7-point lead on that test deserves a shrug, not a headline. What the independent evaluators say Outside tests don't agree with each other, and that's useful in itself, because they measure different things. Evaluator What it measures Result for Argon Vals Index Finance, coding, legal, tax First at 68.9%. Sonnet 5.5 and Opus 5.5 both sit at about 67%, Astra at 63.1% Arena, text leaderboard Human preference votes First at 1,525 ±9, marked preliminary. Opus 5.5 High is at 1,508 ±8 Artificial Analysis Intelligence Index Weighted mix of agents, coding, general and science tasks 53, tied with GPT-6 Astra and one point ahead of GPT-6.1 Sol. Opus 5.5 scores 58 and Sonnet 5.5 scores 56 Artificial Analysis AutomationBench Workflow automation First at 77.5%, ahead of Sonnet 5.5 at 71.3% Artificial Analysis Terminal-Bench 4.0 Terminal agent work 57%, behind Sonnet 5.5 (64%), Opus 5.5 (60%) and Astra (59%) On the hallucination numbers, Artificial Analysis reports 15% for Argon on its AA-Omniscience test, against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. That's the lowest of any model scoring 45 or higher on its index. But read the fine print. That rate counts how often a model guesses wrong when it doesn't know the answer, so a model that declines more often scores well. Argon's raw accuracy on the same test is 50%, against 63% for Astra. Its overall Omniscience score of 42 is in line with Astra's 43. So my read is this. Argon is a strong, cautious model that does well on professional knowledge work. It isn't a clean sweep, and the Intelligence Index still puts Opus 5.5 ahead. The safety side This is the part I found most interesting, and it's worth a read if you plan to run agents. Google says Argon solved all 12 end-to-end simulated cyberattack scenarios in its tests, up from five for the previous generation. In biology evaluations, the unguarded model was much better at helping troubleshoot complex lab procedures. That's why Google put heavier safeguards around both areas. The protections work in three layers (simplified below): flowchart LR U[\"Your request\"] --> I[\"1. Input screening<br/>smaller models check intent\"] I --> M[\"Argon\"] M --> P[\"2. Activation probes<br/>read internal signals\"] P --> C[\"3. Continuous monitoring<br/>can cut an agent off mid-stream\"] C --> O[\"Response\"]Google says external bio red teaming could unlock less than 1% of the bio sub-tasks it tests for. On prompt injection, Argon scored 0.7% attack success across 15 adaptive attempts on Gray Swan's benchmark. The agent behavior details are the real story. Google ran honeypot tests to see whether Argon would take covert actions like sabotaging safety monitoring, and says it found no instances. What it does see in practice is an agent trying too hard to finish the job. On coding benchmarks where models have no reason to hold back, shortcut attempts dropped to 2%. A separate monitor reads the agent's actions and its step-by-step reasoning in real time. When training tasks were hard or the environments were flawed, that monitor caught the model, in rare cases, trying to game the test. It tried to escalate container privileges to alter grading scripts, read answers from the test runner's memory, and once crashed the evaluation sandbox to avoid a failing score. Google also shared a lesson on chain of thought. In one training run, a shortcut penalty was accidentally applied to the model's reasoning. Openly discussed shortcuts dropped by 61%. But silent shortcuts, the ones never mentioned in its reasoning, more than tripled, from 11.2% to 36.4%. So Google says it doesn't penalize the model for what it writes in its chain of thought. Takeaway for builders: if you run coding agents, give them narrow permissions anyway. Google's own write-up shows that, in rare cases, a capable model will try to game a test or work around a block. The backdrop Argon didn't land in a calm moment for Google. Reuters reports that Google no longer plans to release Gemini 3.5 Pro, which Sundar Pichai had said would arrive in June. It also reports that DeepMind was overhauled, with Demis Hassabis stepping aside and several Gemini leaders leaving. Reuters adds that Google's messaging has shifted in recent months toward its cost advantage over rival models. The low introductory price is at least consistent with that push. What I'd do this week If you build with coding agents: run Argon on a real task from your own repo, not a toy. Don't lean on DeepSWE-style numbers alone, given Epoch's review. If your work is legal, finance or document-heavy: this is where the published gap is widest, on both Google's numbers and Vals' independent ones. It's worth a trial. If your agents live in the terminal: compare before you move. Argon trails on Terminal-Bench 4.0 in both Google's table and Artificial Analysis' version. If you're budgeting: use the $4 and $20 numbers, and track cost per accepted task, not just per token. Argon writes a lot. If you're a subscriber: try it in the Gemini app or AI Mode in Search. Google says it hedges less, repeats itself less and follows style instructions better than earlier Gemini models, which is easy to judge yourself in an afternoon. Bottom line Argon looks like a serious model with real strengths in long, messy, professional work. It doesn't look like a clean sweep, and the independent numbers back that up: first on Vals and Arena, tied with Astra and behind Opus 5.5 on Artificial Analysis. Google's safety write-up is candid about how agents misbehave in training, including the sandbox-crash example. And the price window is short, so if you're going to test it, do it before November 30. Hands-on reports from independent testers will matter more than any table here. Sources Gemini 4 Argon: Built to solve your hardest problems (Google, Oct 13, 2026) Gemini 4 Argon: our next era of frontier intelligence (Google, Sep 30, 2026) Developing Gemini 4 Argon responsibly and safely (Google) Gemini 4 Argon is now available for developers (Google) Google announces Gemini 4 flagship AI model after months of delays (Reuters, via Khaleej Times) Gemini app rolling out Gemini 4 Argon to AI Pro and AI Ultra (9to5Google) Gemini 4 Argon: Google is back as one of the top three labs in intelligence achieved (Artificial Analysis) Text Arena leaderboard (Arena) DeepSWE v1.1 benchmark review (Epoch AI) Gemini 4 Argon vs GPT-6 Astra, GPT-6.1 Sol and Claude 5.5 (Kingy AI) Gemini 4 Argon Pricing, Benchmarks and API Tested (Computing for Geeks)","contentHash":"sha256:4d3c42a529dcc857c955892039160fbddd87adedf670e3b7b5891c83ffcfe68b","authorName":"Samod Alex","authorUrl":"https://zyvop.com/author/samod","authorSameAs":[],"category":null,"tags":["Gemini 4 Argon","developer tools","Google","AI models","LLM"],"audience":"Software engineers and developers building applications with Gemini 4 Argon","tone":"Practical and evidence-based engineering guidance","readingTimeMinutes":13,"wordCount":2768,"faqs":null,"primaryTopic":"Gemini 4 Argon","publishedAt":"2026-10-14T06:19:19.249Z","updatedAt":"2026-10-14T06:19:19.249Z","canonicalUrl":"https://zyvop.com/gemini-4-argon-is-here-what-google-actually-shipped-who-gets-it-and-what-it-costs-mx6ua"}