Gemini 4 Argon

Gemini 4 Argon is Google DeepMind’s AI model for difficult software engineering, research and professional tasks that take many steps to complete. It is available through OpenRouter as google/gemini-4-argon, checked on 14 October 2026. The current listing gives a 1,048,576-token context window and standard prices of $2 per million input tokens and $10 per million output tokens.

Google announced Argon on 30 September 2026, with cyber defenders and trusted testers receiving the first access. Its public results cover coding, finance, legal work, science, visual understanding and vulnerability remediation. YFarmX’s new measurement adds a separate finding: Argon’s reported token counts form a distinct fingerprint in our dataset.

Who can use Argon?

Google’s API release notes date general availability to 13 October 2026. The Gemini app announcement gives Google AI Pro and Ultra subscribers access. Copilot is rolling it out to Pro+, Max, Business and Enterprise under its own access rules.

OpenRouter serves Gemini 4 Argon under google/gemini-4-argon on 14 October 2026. Its provider listing covers Google and Google AI Studio routes. The model accepts text, images, files, audio and video, and returns text. Listed reasoning settings are low, medium and high, alongside tools and structured output.

Google’s initial rollout on 30 September 2026 served trusted cyber defenders through the Fairwind Program and trusted testers. Google said those defenders and its internal teams would receive a configuration with cyber guardrails removed. That access programme has its own conditions. Google’s launch announcement explains the original rollout.

Our fingerprint shows different token counts

YFarmX measured Gemini 4 Argon on 14 October 2026 across 50 fixed text samples, covering languages, digits, code, whitespace and emoji. Each sample was measured twice. All 50 readings reproduced, and the fixed overhead stayed stable. Argon’s counts match the earlier Gemini 3.8 Flash record on 28 of 50 samples, or 56%.

The changes vary in size and direction, placing Argon in a distinct token-count group under our comparison method. The eighteen control-marker samples also reproduced across both measurements. Their counts give an inconclusive family attribution. These measurements describe the endpoint’s reported token accounting; claims about model weights or training lineage require additional evidence.

The Argon measurement record carries the full counts and comparisons. The AI Fingerprinting methodology explains how to interpret a shared or distinct fingerprint. The comparison above uses the earlier Gemini 3.8 Flash observation from 5 October 2026.

The output limit grows to one million tokens

Gemini 4 Argon’s announced output limit is 1M tokens, up from the previous 64K. A token is a piece of text handled by the model, such as part of a word. The larger limit gives an extended reasoning and generation run more room to continue through a complex problem.

Google’s current developer guide specifies one 1,048,576-token window shared by input, thinking and output. The default output allowance is 65,536 tokens. Longer generation uses continuation or background execution and remains limited by the space left after the input.

For the current OpenRouter Google route, the provider listing on 14 October 2026 specifies 65,536 output tokens. Google’s one-million-token announcement describes its broader model capability. Choose a service and route whose declared allowance fits the length of the work you want to generate. OpenRouter’s provider settings

Introductory prices run through 30 November

Google’s current price schedule keeps the introductory rates through 30 November 2026. From 1 December, it lists $4 input, $20 output and $0.20 cached input per million tokens.

Google’s announced introductory Argon API prices on 30 September 2026 are $2 per million input tokens and $10 per million output tokens. Cached input receives a 95% discount on the input price, which makes the introductory cached rate $0.10 per million tokens. Caching lets eligible repeated input be reused at the discounted rate.

The launch announcement’s pricing footnote says the input and output rates become $4 and $20 per million tokens after the introductory period. These are announced API terms for the rollout. Google AI Ultra is a separate subscription route named in Google’s access plan. Google’s pricing terms

Coding results depend on the task

Google reports 77.9% on DeepSWE v1.1 for Argon at launch, ahead of GPT-6 Astra and Claude Opus 5.5 in its published comparison. DeepSWE tests long-horizon software engineering. Other coding tests reward different work, and the launch table gives a fuller view of the models’ strengths.

Figures are from Google DeepMind’s launch table. Higher scores indicate better performance within each listed test. Vals Index measures professional work; AutomationBench tests end-to-end business tasks; LVBench tests long-video understanding; CWE-bench v1 tests vulnerability remediation.

Google’s evaluation methodology says Argon generally uses the highest thinking setting and single-attempt scoring, with exceptions explained for individual tests. Google computes its own DeepSWE result using a mini-swe agent harness. Several comparison scores come from providers’ system cards or the benchmark operators’ leaderboards. The table therefore combines results from those stated evaluation sources.

The independent Artificial Analysis listing, checked on 30 September 2026, scores Gemini 4 Argon (High) at 53 on its Intelligence Index. That index is a separate assessment with its own test mix and scale. It gives readers an additional view alongside Google’s task-specific launch results.

Google is using Argon to rewrite and optimise code

Google’s 30 September 2026 announcement describes Argon agents working on C and C++ migrations to Rust, from core libraries to more than 800,000 lines in the Fuchsia Zircon kernel. Google says these critical-system rewrites undergo automated and manual audits, emulation testing and review before production deployment.

For libgav1, Google’s video decoder, agents replaced 32,000 lines of SIMD code in an existing Rust port through repeated profiling and compiler analysis. Google reports that the resulting memory-safe decoder runs 2.7 times faster than that Rust port, with identical video output. The comparison uses the earlier Rust port as its baseline.

Google also reports more than 300 TiB of freed memory from optimisations applied across its data centres. Separately, its quantum researchers used Argon to improve a subroutine’s spacetime-resource requirement by 40% against a published baseline in one example. These are Google’s internal engineering and research results. Google’s announcement

Cyber defenders get the first rollout

Google says Argon can find, validate and patch software vulnerabilities. Its launch table gives the model 68.0% on CWE-bench v1, equal to GPT-6 Astra in that table. Google’s methodology identifies CWE-bench as a public-leaderboard result, scored by single-attempt success, with ties ranked using four-attempt success.

Wiz is already using Argon through Scan for Good, its initiative to find and remediate high-risk exposures affecting public infrastructure. Google reports that Argon identified a critical vulnerability exposing personal information in healthcare software used by hospitals worldwide. The announcement also describes internal vulnerability tests spanning 20 programming languages and Wiz’s tests against web applications with source code withheld from the model. Google

Google is strengthening safeguards before wider access

Google’s Argon rollout pairs tester feedback with work on misuse prevention, prompt-injection resistance, monitoring for misalignment and secure execution environments. Google describes safeguards that monitor internal activations for misuse and systems that monitor the model’s reasoning and actions, stopping execution when needed. Its announcement also describes isolating environments used for high-risk training and evaluations.

For developers choosing a model, the current provider’s price and output allowance are practical starting points. Argon’s published coding results, broad input support and the new YFarmX measurement provide further evidence to compare alongside the demands of the job. OpenRouter’s current listing gives its available settings.