Gemini 4 Argon is now rolling out to everyone — strongest on long, heavy, multimodal work, still last of four on hard terminal coding

Benchmark margins: Argon vs its named comparator

Each row is one benchmark; the circle is Argon, the open circle its comparator, sorted by Argon's effective margin. The Gray Swan row measures attack success rate, where lower is better — its bar colour already reflects that reversal. Scores are not averaged: the benchmarks measure different things on different scales.

Scorecard: 8 wins, 2 ties, 5 losses

Against the named comparator on each of the 15 benchmarks, after reversing the lower-is-better Gray Swan row.

Where it wins and where it loses, by workload

Effective margin in points versus the comparator (Gray Swan reversed so right always means Argon is better).

Launch facts

Who can use it, where, and what it takes

Access map from Google’s 13 October 2026 rollout announcement. Every channel needs a paid tier or an application — there is no free-tier Argon in the checked rollout details. Note the difference at the bottom: everyone else gets Argon with its standard guardrails; only vetted Fairwind Program defenders get the guardrail-free cyber build. blog.google

Where Argon is strong — and where to test before you commit

Workload suitability from 32 benchmark and deployment evidence rows. Green cells are documented leads; amber cells are areas where Argon trailed rivals, so pilot it against alternatives before standardizing. Deployment figures (memory savings, Rust migrations, the quantum result) are Google-reported from internal use. blog.google, thejayant.in

The pattern: Argon is strongest when work is long, context-heavy, multimodal, specialized, and sustained across many steps — inside Google it freed 300+ TiB of memory, migrated C/C++ to Rust up to an 800K+ line kernel, produced a memory-safe libgav1 decoder 2.7× faster than the prior Rust port, and beat a published quantum-subroutine baseline by 40%. For hard terminal coding or autonomous computer operation, it finished last of four more than once.

How to prompt it: the anatomy of a good request

Based on Google’s official guidance for its Gemini reasoning models. Caveat: much of this documentation is written for Gemini 3 / 3.5, not an Argon-only guide — the principles carry, the version specifics may not. ai.google.dev

1. Be precise and direct — state objective, audience, constraints, acceptance criteria and output format; skip persuasion and filler. ai.google.dev
2. Put critical rules, role and output format in the system instruction or at the very beginning of the prompt. ai.google.dev
3. Separate sections consistently with Markdown headings or XML-style tags such as <context> and <task>; use one format per prompt. ai.google.dev
4. For long documents, codebases, datasets or videos: give all context first, put the question at the end, and start it with “Based on the preceding information…”. ai.google.dev
5. For grounded research: instruct it to use only supplied sources, cite every material claim, distinguish fact / inference / unknown, and say when evidence is missing. ai.google.dev
6. Match thinking effort to difficulty: high for complex coding, maths or multi-step planning; default for moderate work; low for retrieval or classification. ai.google.dev
7. To cut cost, lower the thinking level instead of shrinking max output tokens — a hard token cap can truncate mid-reasoning while still billing thinking tokens. ai.google.dev
8. Preserve the full conversation history for iterative debugging and refactoring; carried reasoning context improves multi-turn work, at a token cost. ai.google.dev
9. For tool-using agents, impose an explicit tool-call budget and require verification before irreversible actions. ai.google.dev
10. Ask for the verbosity and tone you want — Gemini defaults to direct, efficient answers. And skip long visible chain-of-thought: ask for a concise rationale, assumptions, checks and the final answer. ai.google.dev

Five copyable templates

Drop-in starting points built on the guidance above; replace the bracketed slots.

Read the numbers with care

All 15 benchmark rows