Gemini 4 Argon: Specs, Context Window, Pricing, Benchmarks and Availability

Gemini 4 Argon is a proprietary large language model from Google, generally available since 13 October 2026 under the API identifier gemini-4-argon. Google announced it on 30 September 2026 and calls it its frontier model for complex work in coding, finance, legal, cybersecurity and science. This profile collects the published facts on one page: specifications, context and output limits, prices, where the model runs, restricted variants, provider-reported benchmarks, data handling and retirement status. It is written for developers, architects and buyers who need the numbers with their sources. Every fact comes from Google's own documentation, pricing pages, announcement and model pages.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Generally available since 13 October 2026. Gemini 4 Argon reached GA on the Gemini API and on Google Cloud's Gemini Enterprise Agent Platform on that date, after a 30 September 2026 announcement.
• One shared 1,048,576-token window. Input, thinking and output share one limit; a single segment returns up to 65,536 tokens streamed, and long decode extends output towards the full window.
• Prices double on 1 December 2026. Standard pricing is $2 input and $10 output per million tokens until 30 November 2026, then $4 and $20; the model is not on the Gemini API Free tier.
• A cyber permissive variant is allowlist only. gemini-4-argon-cyber-permissive is for defensive security work and is granted through Google's Fairwind Program.
• No model card or knowledge cutoff is published. Google's model pages and developer guide carry the specifications; neither states a training cutoff or links a model card.

What is Gemini 4 Argon?

Gemini 4 Argon is the first model in Google's Gemini 4 family, according to its Google Cloud developer guide. The Google Cloud model page describes it as "Google's frontier model for complex work in coding, finance, legal, cybersecurity, and science", and the Gemini API changelog entry for 13 October 2026 calls it "our most capable model built for complex coding, agentic, math, and science tasks". Google's announcement blog post of 30 September 2026 positions it for deep reasoning in long-running workflows. The family later added Gemini 4 Carbon and Gemini 4 Hydrogen; see the Google Gemini Models hub for the full catalogue.

Argon is a thinking model. It accepts text, images, audio, video and PDF documents, and returns text. Google publishes no parameter count and no open weights. Its developer guide changes several API conventions compared with earlier Gemini models, covered under Modalities and tools.

Key facts at a glance

As of 2026-11-15.

Release and version history

The announcement said access would roll out gradually, including a voluntary United States government pre-release access process, and would reach paid API customers and Google AI Ultra subscribers first. Google's developer guide describes only one migration path into Argon: from gemini-3.1-pro-preview. Google has published no dated point releases or snapshot identifiers for Argon on the pages checked.

How much does Gemini 4 Argon cost?

As of 2026-11-15, USD per million tokens, from the Gemini API pricing page. Output prices include thinking tokens. The first figure applies until 30 November 2026, the second from 1 December 2026.

• Caching: implicit only. Explicit context caching is not supported, so there is no separate cache storage charge listed for Argon.
• Long context: the pricing page lists no separate long-prompt tier for Argon.
• Grounding: Grounding with Google Search and with Google Maps each include 5,000 free requests a month, shared across Gemini 3 and newer models, then $14 per 1,000.
• Long decode: the developer guide states that continuation segments that hit the implicit cache receive a 95% caching discount, and that every processed token, including thinking, is billed.
• Google Cloud: on the Gemini Enterprise Agent Platform, Argon is sold through Standard PayGo with Usage Tiers, Priority PayGo, batch inference and Provisioned Throughput; fixed quota is not offered. Google Cloud pricing is published on its own pricing page.
• Gemini app: Google's subscriptions page lists Argon in Google AI Pro ($19.99 a month) and Google AI Ultra (from $99.99 a month).

Worked cost examples

Illustrative arithmetic from the published rates above, before grounding charges and without caching. Actual bills depend on thinking tokens, which count as output.

For prices across providers see Frontier AI Models Compared.

Context window and output limits

Google's announcement states that maximum output rose to 1 million tokens from 64,000. Long decode returns status values requires_continuation, incomplete and completed, and the guide notes that it needs the v1beta1 API. Google does not recommend long decode for chat, classification, latency-sensitive calls or prompts near the input limit. Setting max_output_tokens above 1,048,576 returns HTTP 400. See Context Windows Explained.

Modalities and tools

The developer guide lists API conventions that differ from earlier Gemini models:

• Thinking control: thinking_budget and thinking_level="MINIMAL" return HTTP 400; use thinking_level. Google suggests LOW for classification and extraction, MEDIUM for general coding and document questions, HIGH for multi-file engineering and proofs.
• Sampling: custom temperature, top_p and top_k values are ignored; frequency_penalty, presence_penalty and candidate_count return HTTP 400.
• Tool turns: function responses must match the preceding call's id, name and order.
• Chat history: histories cannot end with a model turn, and prefilled model turns are unsupported.

See Reasoning Models Explained and, for agents that drive screens, Computer-Use and Browser Agents.

Where is Gemini 4 Argon available?

ML processing for the regional endpoints runs in us, eu, europe-west2 and asia-southeast1. For security set-up of the developer channel see the Gemini Enterprise Security Guide.

Access restrictions

• Standard model: generally available to paid Gemini API customers and Google Cloud customers; no allowlist.
• gemini-4-argon-cyber-permissive: allowlist only, through the Fairwind Program or a customer's Google team. Google describes it as a Google Frontier Cyber Model subject to Section 31 of the Service Specific Terms, intended for defensive work such as vulnerability discovery, reverse engineering, exploit verification and patch generation.
• Cyber variant differences: default thinking level HIGH; computer use not supported; regions global, us and eu only. Used with CodeMender it falls under the Pre-GA Offerings Terms. Access to Gemini 3.8 Flash Cyber does not carry over and needs a separate application.
• Announcement statement: Google's blog says it planned to release Argon without cyber safeguards for trusted defenders and its internal teams.

Benchmarks (provider-reported, dated)

All scores below are reported by Google. Sources: Google blog announcement (2026-09-30) and the Google DeepMind Gemini model page, whose table is undated (as shown on 2026-11-15). Benchmark versions and settings follow the provider's wording; see How to Read LLM Benchmarks.

The DeepMind page also lists Agents' Last Exam v1.1 (69.0%), RiemannBench (76.0%), Chartography (71.6%) and ExploitGym (42.0%). Google's blog adds that Argon leads the Gray Swan indirect prompt injection benchmark; no score is given there.

Model card and system card

Not stated by Google on the pages checked (2026-11-15): no document labelled model card or system card is linked from the Google Cloud model page, the developer guide, the AI Studio models page or the DeepMind Gemini page. The Google Cloud model page and developer guide are the specification documents. The DeepMind page lists an October 2026 post titled "Developing Gemini 4 Argon responsibly and safely". The announcement names four safety areas: misuse protection guided by Google's Frontier Safety Framework, including refusals for CBRN and cyberattack requests; prompt injection resistance; monitoring of the model's reasoning and actions for misalignment, with the ability to halt execution; and hardened sandboxes for training and evaluation. See Model Cards and System Cards Explained.

See the Frontier AI Model Risk Frameworks Guide for how such reports are structured.

Data handling

• Gemini API paid tiers: the pricing page states that Standard, Batch, Flex and Priority content is not used to improve Google's products.
• Contributor tier: content in this invite-only preview tier is used to improve Google's products.
• Google Cloud controls: the model page lists data residency, customer-managed encryption keys (CMEK), VPC Service Controls and Access Transparency (AXT) for online prediction, batch inference and context caching.
• Implicit caching: durable caching is on by default for shared prompt prefixes of 4,096 tokens or more.
• Long decode storage: stored continuation segments are kept for 7 days.

For Gemini developer tooling see the Gemini CLI and Gemini Code Assist Security Guide and Zero Data Retention for AI.

Deprecation status

Gemini 4 Argon is current and generally available. Google Cloud lists a retirement date of 13 October 2027 or later and states that models may remain accessible through the Gemini API after that date. No deprecation notice for Argon appears in the Gemini API changelog as of 2026-11-15. Retirement dates can move, so check the model versions page before planning. See the AI Model Retirement Guide.

How Gemini 4 Argon differs from other models in its family

As of 2026-11-15. Models listed alphabetically. Sources: Google Cloud model versions page and Google's Gemini subscriptions page.

• Lifecycle table: Google Cloud lists Argon in its generally available table, with a retirement date; Carbon and Hydrogen sit in the table for models with shorter availability periods.
• Order of release: Argon was the first Gemini 4 model, released four weeks before Carbon.
• Regions: the Argon model page lists global, us, eu, europe-west2 and asia-southeast1; check each sibling's own model page for its regions.
• Cyber variant: Argon has an allowlisted cyber permissive version; see the hub for the other models' variants.

See the Gemini 4 Carbon profile and the Google Gemini Models hub.

Practitioner checklist

• Re-estimate budgets for the 1 December 2026 price change on every Argon tier.
• Confirm the project is on a paid Gemini API tier, or a Google Cloud PayGo or Provisioned Throughput plan.
• Replace thinking_budget with thinking_level and remove sampling and penalty parameters before migrating from gemini-3.1-pro-preview.
• Remove explicit caching code and design prompts with stable prefixes of 4,096 tokens or more for implicit caching.
• Use long decode only for long generation jobs, on the v1beta1 API, and plan for 20-minute segments.
• Check that the required Google Cloud region offers the consumption option you need.
• Apply through the Fairwind Program only for defensive security use cases that need the cyber permissive variant.
• Store Gemini API keys in a secrets manager and restrict them by API and project.

Standards and references

• Google Cloud: Gemini 4 Argon model page (accessed 2026-11-15)
• Google Cloud: Gemini 4 Argon developer guide (accessed 2026-11-15)
• Google AI Studio: Gemini API pricing (accessed 2026-11-15)
• Google AI Studio: Gemini API changelog (accessed 2026-11-15)
• Google AI Studio: Models (accessed 2026-11-15)
• Google blog: Gemini 4 Argon announcement (accessed 2026-11-15)
• Google DeepMind: Gemini models (accessed 2026-11-15)
• Google: Gemini subscriptions (accessed 2026-11-15)

Related NHI Mgmt Group resources: Google Gemini Models, Gemini 4 Carbon, Frontier AI Models Compared, Gemini Enterprise Security Guide.