Gemini 4 Argon: Built to solve your hardest problems

What it claims

Gemini 4 Argon is Google DeepMind's new frontier model, aimed at long, multi-step work in coding, finance, legal, cybersecurity and science. The launch price of $2/M input and $10/M output is labelled introductory, and cached input tokens are 95% off. Most of the evidence is internal case studies rather than benchmark tables. Several of the results are public code you can inspect: ADK optimizations are merged in google/adk-python, xTB speedups are contributed to the official xTB repository, and the vLLM changes for Gemma 4 E2B are upstream pull requests.

Evidence

Engineering case studies

• Google-internal Antigravity use: agentic code submissions rose more than 35% and code rollbacks fell nearly a third. No time window or baseline population is given.
• ADK: Argon worked overnight from an .agents/skills rulebook and a one-paragraph goal. The result was 40% less in-memory session prep time in long sessions and 21% faster cold imports.
• xTB: up to 2.3x faster molecular structure prediction and 1.8x faster conformer search.
• Gemma 4 E2B on vLLM, single A100: throughput went from about 37,900 to over 108,000 tokens/s, a 2.85x end-to-end speedup. The page reports 100% token parity. Changes span attention tuning to speculative decoding, and the work took "a few days."
• Chromium: Argon helped separate a ~6M-line legacy feature folder within the 40M-line codebase, with refactors across hundreds of files.

Science case studies

• AlphaQubit: an internal version was translated into a single optimized kernel, with inference throughput roughly doubled.
• Quantum arcsin circuit: the version is formally verified in Lean and cuts spacetime cost by more than 100x against a published quantum CORDIC baseline.
• Gamow Labs private evals: RareBench 57% and LabBench 52.2%, both described as "topped." Competitor scores are not given.
• Enzyme design: five novel serine esterase designs against 4MU-acetate, with lab-confirmed catalytic activity.

Benchmarks

• ERQA: 87%, "relying on pure reasoning." This means no tools, and no competitor numbers are shown.
• Vals Index: ranked #1, but no score is given.

Safety

• Bio red teaming: external red teaming unlocked fewer than 1% of the bio sub-tasks tested.
• Adversarial violations: reduced nearly 2x versus previous Gemini models.
• Shortcut attempts: 2% on coding benchmarks.
• Gray Swan IPI: 0.7% attack success across 15 adaptive attempts per attack.
• Misalignment monitor: it reads reasoning and actions. It was used during RL training without penalizing the model's thoughts, so the reasoning stays monitorable.
• Trained behaviour: Argon is trained to pause for human confirmation before irreversible steps. Agent builders should expect these pauses.

How to use it

• Developers: the Gemini API paid tier is required. AI Studio, Antigravity and Android Studio access is listed for Google AI Ultra subscribers.
• Enterprises: Gemini for business, the API, and Antigravity.
• Context: the 1M limit is unified input+output, not 1M input plus a separate output budget. Long outputs use a new continuation protocol. The page points to the developer docs for setup but does not describe the API shape.
• Prompting tips mentioned: Antigravity slash commands such as /goal and multi-agent loops with custom skills.
• Researchers: 10,000 scientists can get Antigravity's highest usage tier free for 6 months. Applications are open, subject to eligibility, capacity and selection.
• Security teams: the Fairwind Program gives access to Argon without cyber guardrails. It is expanding by hundreds of partners, with rolling admission.

Limits

• The pricing is explicitly introductory, so budget for it to change.
• Nearly all gains are self-reported case studies run by Google or early-access partners, with no variance or failure counts.
• The RareBench and LabBench evals are private.
• Several "outperforms" claims (Vals, ERQA, RareBench, LabBench) give no competitor scores.
• The cornhole video-to-sim-to-real demo is a single internal trial (seven consecutive shots on Apptronik Apollo).
• Default API access carries cyber guardrails, and unrestricted use requires Fairwind admission.
• Detailed safety evaluation is deferred to the model card, which is not included here.