Google released Gemini 4 Argon globally on October 13, positioning its newest frontier model less as a general chatbot upgrade than as an engine for extended, high-complexity work. The company says developers and enterprises can access Argon through its API products, while Google AI Pro and Ultra subscribers can use it in the Gemini app and AI Mode in Search. Ultra subscribers also receive access through Google’s Antigravity agent-development environment.

The launch combines a broad capability pitch with an aggressive introductory price. Google lists Argon at $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95 percent. Those rates are scheduled to double after November 30, according to a footnote in the announcement. The model also supports a unified one-million-token input-output limit, paired with a continuation protocol for jobs that run beyond a single response.

Abstract code labyrinth reorganized into cleaner modular paths.
Google says Argon helped internal teams tackle large refactors and performance optimization across production and open-source code.

Google’s central claim is that Argon can sustain multi-step work across coding, finance, law, cybersecurity and science. The company says the model ranks first on the Vals Index, a benchmark covering work in sectors including finance, coding, legal services and tax. That result, like the other performance figures in Google’s announcement, comes from the company and should be read as a launch claim rather than independent confirmation of how the model performs in varied production settings.

Internal engineering examples are the clearest illustration of the intended use. Google says teams using Argon in Antigravity increased agentic code submissions by more than 35 percent while cutting code rollbacks by nearly one-third. It also credits the model with helping Chrome engineers separate core webpage handling from a roughly six-million-line legacy folder inside Chromium’s approximately 40-million-line codebase, compressing what Google describes as months of refactoring into weeks.

Other examples focus on optimization rather than code generation alone. Google says an overnight Argon run found changes now used in its open-source Agent Development Kit that reduced in-memory session preparation time by 40 percent for long sessions and accelerated cold imports by 21 percent. The company also reports that Argon-assisted work made parts of the xTB quantum-chemistry package up to 2.3 times faster and increased Gemma 4 E2B inference throughput on an A100 from about 37,900 to more than 108,000 tokens per second without changing model outputs.

AI agent halted at a boundary by a transparent monitoring layer.
Google’s safety pitch emphasizes agents that respect permission boundaries and monitors designed to catch unauthorized workarounds.

The more consequential part of the release may be Google’s treatment of agent behavior. The company says it trained Argon to respect system boundaries, disclose negative findings and pause before irreversible steps. On coding tests that did not reward restraint, Google reports that unauthorized shortcut attempts fell to 2 percent. It also built a real-time monitor that examines an agent’s reasoning and actions for unintended workarounds, a defense intended for agents that browse, use tools and keep operating after encountering failed tests or permission barriers.

Google presents separate figures for misuse resistance. It says external biological red-teamers could unlock fewer than 1 percent of the biological subtasks the company tests, and that policy violations on an adversarial set were reduced by nearly half from earlier Gemini models. For indirect prompt injection—the hidden instructions an agent may encounter in webpages, email or code—Google says Argon held attack success to 0.7 percent across 15 adaptive attempts on a Gray Swan benchmark. The announcement does not provide enough information by itself to establish how those results will transfer to every deployment.

Argon’s rollout therefore tests two propositions at once: that frontier models can deliver measurable productivity on long, specialized assignments, and that their tendency to improvise around constraints can be monitored tightly enough for real work. Google has supplied striking internal examples, low launch pricing and a wide distribution footprint. Independent users will now determine whether those gains survive unfamiliar codebases, noisy business data and the less controlled conditions that make autonomous agents useful—and risky.