Gemini 4 Argon: Built to solve your hardest problems

Google DeepMind has launched Gemini 4 Argon, a new frontier model it describes as its most capable yet for complex, multi-step work. The company said the model is globally available starting today for developers and enterprises through the API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect at Google, announced the release.
The introductory price is $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%. Argon supports a 1 million token unified input-output limit through the Gemini API, which Google calls industry-leading and which matters for agentic workloads that accumulate long histories of code, documents and tool calls.
Google says Argon excels at deep work across coding, finance, legal, cybersecurity and science, and that its depth and precision on multi-step tasks let it outperform other frontier models on benchmarks spanning those domains. It ranks first on the Vals Index, which measures model performance across economic sectors including finance, coding, legal and tax work. On the Embodied Reasoning benchmark ERQA, which tests physical and spatial reasoning in 3D environments, Argon scores 87%, ahead of other frontier models according to Google. On Gamow Labs evaluations derived from real clinical and lab data, it scored 57% on RareBench rare-disease diagnosis assistance and 52.2% on LabBench wet-lab reasoning.
Internally, Google reports that Argon in Antigravity increased agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. The Chrome team used it to separate core webpage handling from a roughly 6 million line legacy folder inside Chromium's 40 million line codebase, completing an architectural refactor across hundreds of files in weeks rather than months. The change is intended to let live webpages move between tabs, side panels and floating windows without reloading, and to eliminate a class of browser crashes. On Google's open-source Agent Development Kit for Python, Argon worked overnight from a rulebook and a one-paragraph goal and produced optimizations now live in google/adk-python that cut in-memory session preparation time by 40% in long sessions and sped up cold imports by 21%.
The model also contributed to open scientific software. Google researchers used Argon on xTB, a widely cited semiempirical quantum chemistry package, reaching up to 2.3x faster molecular structure prediction and 1.8x faster conformer search, with changes contributed upstream to the official GitHub repository. On Google's open Gemma models, Argon helped achieve a 2.85x end-to-end inference speedup on an A100 via vLLM, raising throughput from about 37,900 to over 108,000 tokens per second on Gemma 4 E2B with full token parity and no quality loss, a task Google says would normally take months.
Google highlighted visual and 3D work as an area of improvement over earlier Gemini models, citing better spatial reasoning, mathematical precision and visual coding. Examples include a procedurally modeled Pagoda built through multi-agent visual critique loops in Antigravity, a playable miniature racing game generated from photos of a real desk using Blender MCP, Nano Banana textures and Three.js, and a pure-Python 3D ray-tracing engine produced without external libraries. Enterprise and research users are cited as well: Isomorphic Labs is using Argon to synthesize complex PDFs and patents, work Google says typically takes weeks but can now take hours, and Gamow Labs used it to identify a chromosomal rearrangement that routine software had missed.
On scientific research, Google reports Argon translated an internal version of its AlphaQubit quantum error correction decoder into a single optimized hardware kernel, nearly doubling inference throughput, and wrote a Lean-verified implementation of the arcsin quantum subroutine that cuts spacetime cost by more than 100x versus a published CORDIC baseline. In fusion, a Google team working with Professor Matt Landreman of the University of Maryland found a broader class of counterexamples to Grad's conjecture about stellarator magnetic fields, giving the community new benchmarks. In enzyme design, Argon proposed five novel serine esterase designs against the benchmark molecule 4MU-acetate within hours, and laboratory testing confirmed catalytic activity across them.
The release includes a safety program. Google says external bio red teaming could unlock less than 1% of tested bio sub-tasks, and that on an adversarial set violations fell nearly twofold versus previous Gemini models. Argon was trained to respect system boundaries, report negative findings honestly and pause before irreversible actions, with shortcut attempts on coding benchmarks dropping to 2%. An automated misalignment monitor reads agent reasoning and actions to block unintended workarounds in real time, and on the Gray Swan Indirect Prompt Injection benchmark attack success across 15 adaptive attempts fell to 0.7%.
Google is offering 10,000 scientists the highest usage tier in Antigravity free for six months, onboarded on a rolling basis subject to eligibility, capacity and selection, with applications open now. It is also expanding its Fairwind Program to hundreds of additional partners; Wiz, an early access user, reported using Argon to find a public file download bypass exposing a master key in healthcare infrastructure and to help address a sensitive-data gap with a federal agency.
Why it matters: Developers and product teams gain a broadly available frontier model at a relatively low introductory token price with a 1M-token context window, which favors long-horizon agentic coding and document-analysis workflows where context length drives architecture. The benchmark and pricing claims come from Google itself, so teams evaluating Argon against competing frontier models should verify results on their own workloads before committing. The free Antigravity tier for scientists, if eligibility holds, is likely to matter most for research groups that could not otherwise afford high-volume agentic runs.
Based on reporting from the original publisher. Visit the source for full context and later updates.