Gemini 4 Argon

Gemini 4 Starts With a Frontier Reasoning Model

Google announced Gemini 4 Argon on September 30, 2026 through a phased release to trusted cyber defenders.

The model became generally available through the Gemini API on October 13, 2026.

Its stable API identifier is gemini-4-argon.

Google describes Argon as its latest and most capable model.

The intended workload is not lightweight chat. Argon is designed for tasks that may require extended reasoning, many tool calls, difficult software-engineering decisions, deep mathematical work, scientific investigation, long documents, multimodal evidence, complex business processes, autonomous multi-step execution, and cybersecurity defense.

The model accepts text, images, video, audio, and PDF documents and generates text output.

Argon supports thinking, function calling, structured outputs, code execution, computer use, file search, Google Search grounding, Google Maps grounding, URL context, agentic video understanding, implicit caching, Batch API, Flex inference, and Priority inference.

It does not support native audio generation, native image generation, or the Live API.

The Context Window Is Shared

This is a major conceptual change from earlier Gemini models.

The full budget is 1,048,576 tokens and includes:

• initial input;
• multimodal representations;
• thinking;
• generated output.

A very large input therefore leaves less room for reasoning and generation.

The standard synchronous output limit is 65,536 tokens.

For longer tasks, Argon supports long decode.

Long decode can continue generation in multiple 64K hops while keeping the work inside one interaction.

On the Interactions API, long decode uses background execution. The server generates one hop, continues from that state, adds another model-output step, and repeats until completion or the shared budget is exhausted.

This is useful for long mathematical derivations, extended scientific reasoning, large code transformations, complex research reports, and exhaustive technical analysis.

Google highlights finance, legal work, tax-related reasoning, scientific research, and enterprise process automation.

Argon leads the Vals Index in Google’s launch material, covering economically important tasks across finance, coding, legal, and tax work.

Google also cites strong results on Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.

These workloads are structurally different from generic question answering.

A financial workflow may require source review, calculations, current-data grounding, and a structured conclusion.

A legal workflow may require long-document review, issue identification, research, drafting, and revision.

Argon’s unified context and long decode are directly relevant because both input and reasoning can become large.

Multimodal evidence can stay inside one reasoning workflow.

Argon can combine written instructions, screenshots, charts, diagrams, long videos, audio, and PDF documents.

This is useful for screenshot debugging, financial chart analysis, PDF synthesis, video evidence review, multimodal research, and document-heavy enterprise workflows.

Google reports 91.7% on LVBench, a long-video understanding benchmark.

Argon also supports agentic video understanding; the official model page lists static as the default mode.

Argon does not generate images or audio.

Thinking lets applications trade among reasoning depth, latency, token use, and cost.

Low is useful when the task still needs reasoning but speed matters.

Medium is Google’s default and a practical baseline.

High is appropriate for difficult coding, scientific, mathematical, investigative, security, or long-horizon agentic work.

Thinking tokens consume part of the same 1,048,576-token budget as input and output.

Google recommends raising explicit output caps below 32,768 because Argon may spend more of the allowance on thinking.

Current introductory rates:

• input: $2.00 per 1M tokens;
• implicit cached input: $0.10 per 1M tokens on cache hits;
• output including thinking: $10.00 per 1M tokens.

Beginning December 1, 2026, Google lists:

• input: $4.00 per 1M tokens;
• implicit cached input: $0.20 per 1M tokens;
• output including thinking: $20.00 per 1M tokens.

The cached rate reflects a 95% discount when implicit caching hits.

Argon is not available on the Gemini API Free tier.

For Argon, cost per completed task is more informative than cost per request because reasoning and long decode can materially increase generated tokens.

Google tells developers to remove:

• temperature;
• top_p;
• top_k.

Those sampling parameters are ignored.

Use explicit instructions, structured outputs, and thinking_level instead.

Modern Gemini API behavior also disallows ending a request with a non-empty prefilled model turn.

The Interactions API can preserve multi-turn state through previous_interaction_id, which is useful for function-calling and long-running workflows.

Hard Argon prompts can take several minutes.

For very long generation:

• use background execution;
• set an appropriate max_output_tokens;
• increase client timeouts;
• continue or poll the interaction as required.