Google Releases Gemini 4 Argon for Long-Running AI Agent Work

Google has released Gemini 4 Argon, its newest frontier AI model, to developers, businesses, and subscribers worldwide. The announcement, published October 13, 2026 by Koray Kavukcuoglu, marks the model’s move from limited trusted-tester access to full global availability across the Gemini API, AI Studio, the Gemini platform, Google Antigravity, the Gemini app, and AI Mode in Search.

Argon is aimed at long-running agent work: workflows that plan, use tools, evaluate, and execute over many steps, such as migrating a codebase, running enterprise research, or working through complex cybersecurity tasks. Google reports that the model ranks first on the Vals Index, a benchmark it describes as measuring performance across finance, coding, legal, and tax work. All benchmark figures cited here are Google’s own, unless noted otherwise.

Background: from defenders to everyone

Google first announced Argon on September 30, 2026, with initial access limited to trusted cybersecurity defenders through its Fairwind Program. Logan Kilpatrick of Google said the model then went to Fairwind defenders and early-access customers over the following two weeks, and that more than 100,000 people inside Google tested it over the preceding month. The October 13 announcement extends access to paid API developers, enterprises, and Google AI Pro and Ultra subscribers.

Key details

The defining technical feature is a unified 1,048,576-token context window shared by input and output. Google’s API documentation notes that the default synchronous output is capped at 65,536 tokens, with a new continuation protocol that lets longer generations stay within the shared million-token budget. The model accepts text, images, video, audio, and PDF files as input but produces only text output.

Pricing is introductory and time-boxed. The API launch rate is 2 US dollars per million input tokens and 10 dollars per million output tokens, with cached input tokens discounted 95 percent. Google’s footnote states that after the introductory period expires on November 30, 2026, standard pricing of 4 dollars per million input tokens and 20 dollars per million output tokens applies. For comparison, Cryptopolitan reports that Anthropic’s Claude Opus 5.5 charges 4 and 20 dollars per million input and output tokens respectively, while OpenAI’s GPT-6 Astra charges 10 and 50 dollars.

Access varies by product. Developers use the paid tier of the Gemini API, and enterprises can use the Gemini platform and Gemini for business. Google AI Pro and Ultra subscribers get Argon in the Gemini app and AI Mode in Search, while Google Antigravity, the company’s agent-first development environment, is available to Ultra subscribers. In the Gemini app, a new task mode lets the assistant continue working when the user is away, and the app now connects to more than 40 third-party applications, including Instacart, OpenTable, Adobe, and Shopify.

Google described internal deployments. The Chrome team used Argon in Antigravity to separate core webpage-handling code from a legacy folder within Chromium’s 40-million-line codebase, completing a refactor across hundreds of files that the team said would have taken months. The changes are expected to let live webpages move between tabs, side panels, and floating windows without reloading. Google’s Agent Development Kit team reported that Argon optimized their Python codebase overnight, cutting in-memory session preparation time by 40 percent in long sessions and speeding up cold imports by 21 percent, with the changes now live in the google/adk-python repository.

One outside data point comes from Box. CEO Aaron Levie posted that Box tested Argon on its Box AI Complex Work evaluation for a week and found it scored 14 percentage points above Gemini 3.1 Pro across all its tests, with gains by industry including 37 points in technology and 33 points in consumer products. These are Box’s own evaluation results, not Google’s figures.

On safety, Google said Argon achieved a 0.7 percent attack success rate across 15 adaptive attempts on the Gray Swan indirect prompt injection benchmark. The company is expanding the Fairwind Program, which gives trusted cybersecurity defenders access to Argon without its cyber guardrails, to hundreds of additional partners. Google also said it will offer 10,000 scientists the highest usage tier in Google Antigravity free for six months, with applications opening October 13, 2026. The model card published by Google DeepMind acknowledges hallucinations, possible slowness or timeouts, a March 2026 knowledge cutoff, and describes Argon as a sparse mixture-of-experts transformer.

Why it matters

The launch shows where the frontier model race is moving: away from one-shot chat and toward sustained, tool-using agent work, where the model holds an objective across dozens or hundreds of operations. A unified one-million-token budget plus a continuation protocol is infrastructure for that shift.

The pricing is the other signal. At 2 and 10 dollars per million tokens, Google is undercutting the other frontier labs during a fixed seven-week window, an offer aimed at getting developers to move workloads onto Argon now and measure cost later. That makes the next seven weeks the cheapest window to test whether long-horizon agent work actually reduces labor or merely shifts it into human review, which remains the open question for every agent platform this year.

Sources

• Google Launches Gemini 4 Argon AI Model for Enterprise and Developer Use, Blockchain.News, October 14, 2026.
• Gemini 4 Argon Goes Global at $2/$10, Doubling After Nov 30, FourWeekMBA, October 14, 2026.
• Google launches Gemini 4 Argon, undercutting OpenAI and Anthropic on price, Cryptopolitan, October 13, 2026.
• Gemini 4 Argon: Features, Benchmarks, Pricing and GPT-6 Comparison, InnoAI, October 13, 2026.