Gemini 4 Argon arrives with video understanding, Canvas audio tools and 1M-token agents

What rolled out

Google has made Gemini 4 Argon generally available, a multimodal model that accepts text, images, audio and video as input. According to a dev.to writeup of Google's announcement, developers can reach the model worldwide through the Gemini API, Google AI Studio, Google Antigravity and selected options in the Gemini app. Google positions Argon as a frontier model for long-horizon, high-stakes work in coding, finance, legal, cybersecurity and science, and emphasises agentic video understanding — interpreting footage as part of multi-step workflows rather than simply describing clips.

The rollout also extends Gemini Canvas, an environment for generating apps, games and audio assets. Canvas ships with a built-in Music Synthesizer, which Google's materials present alongside demos of multimedia pipelines that include video tooling and audio mixing.

Context, output length and pricing

Two dev.to posts published hours apart on October 13 cover the launch from different angles, and both report a one-million-token input window through the API. The more technical deep dive adds that output is also capped at one million tokens, a jump from the 16K–64K output ceilings the author says force conventional coding assistants to emit truncated diffs during large refactors.

On pricing, the launch material cited by both posts lists introductory public API rates of $2 per million input tokens and $10 per million output tokens, with a price change flagged after November 30, 2026 that neither post quantifies. The deep dive also reports a 95% discount on cached input tokens, at $0.10 per million — a rate that matters for agents that re-read a large codebase across many turns.

Agentic coding capabilities

The deep dive focuses on software engineering and reports three headline results: a 77.9% state-of-the-art score on the DeepSWE v1.1 benchmark, a 40% end-to-end speedup on session preparation in Google's adk-python toolkit after the model identified redundant per-turn serialisation overhead, and a 0.7% attack success rate against the model on the Gray Swan indirect prompt injection benchmark.

Under the hood, the author describes an Interactions API that keeps an agent stateful across tool calls, carrying the model's reasoning between turns instead of discarding it, with cryptographic signature verification on emitted thoughts. Combined with strict JSON Schema structured output, this is what makes the model usable as a CI gate: audit findings arrive as machine-parseable objects with CWE identifiers, exploit chains and complexity estimates rather than free-form Markdown. The one-million-token output ceiling is the other half of that story, since a whole-module rewrite can come back in a single valid response instead of a partial diff.

Canvas audio and the undocumented feature

The more cautious of the two posts draws a line between what Google documents and what circulates as demo hype. Canvas is a plausible home for audio-centric interactive experiences, but Google's first-party materials do not describe a real-time feature that turns a live webcam feed into a playable synthesizer. Latency, supported camera inputs, how video signals would map to musical controls and the amount of custom engineering required are all unspecified. Teams can build such an experiment from the documented parts — video understanding on the model side, the synthesizer on the Canvas side — but should not plan a product or campaign around it as a packaged capability.

Why it matters

Argon's significance is less any single feature than the combination: video input, a million tokens in and out, cheap cached context and stateful multi-turn APIs together change what an agent can attempt on a real repository or a long video. For developers, the practical move is to test the documented capabilities now — multimodal analysis, long-output refactors, structured audits — while budgeting for the post-November price change and treating viral demos as experiments until Google specifies them.