Gemini 4 Argon

What it claims

Gemini 4 Argon is Google DeepMind's frontier model for long-horizon agentic work. The page backs this mostly with named deployments and partner-reported numbers. The headline engineering case is Chromium. Argon helped the Chrome team separate core webpage handling from a ~6 million-line legacy folder inside the 40 million-line codebase. It submitted multiple code changes at a time across hundreds of files. The page says this compressed "months of work in weeks" and will let live pages move between tabs, side panels and floating windows without reloading.

Evidence

Beyond the CWE-Bench and Box figures already in the summary, the page reports these partner results. All are self-reported, and most do not name a baseline:

• Clinical charting (partner unnamed): on 200+ synthetic longitudinal charts, it had the highest recall and ICD-10 code specificity among the models that partner evaluated, with no external lookups. Validation on real patient records is still in progress.
• Kepler (finance): it matched analyst-built reference models on more than 99% of figures. The tasks include multi-year financial statements built from SEC filings and valuations of companies with complex capital structures.
• Qualia (autonomous research agent): frontier-level on internal multi-step research benchmarks. It finished evals 1.5× faster than "peer models" at a quarter of the cost per task.
• 3D spatial reconstruction (partner unnamed): nearly 100× lower runtime and 90% fewer tokens, versus every model they tested, for turning video into physics-ready 3D scenes.
• Long-running tool-use workflows (partner unnamed): 100% task completion in their evals.
• Roboflow: first in their object-detection evals. It also led their new text benchmark of real-world manufacturing, logistics and aerial images. Argon is now fully integrated into Roboflow.
• Legal (LexisNexis and one unnamed firm): qualitative only. In one evaluation it surfaced a detail the human expert's answer key missed.
• Wiz: used Argon in its free "Scan for Good" programme on healthcare infrastructure. No metric is given.

The 1M-token output window is described as enough to finish reasoning trajectories of 100K+ tokens. It is also pitched for single-pass multi-file app generation and for legacy or schema translation that preserves cross-file dependencies.

How to use it

The page names these access routes but gives no model ID, pricing, rate limits or API parameters:

• Consumer: Google AI Pro and Ultra subscribers can use it today.
• Developers: Gemini API, Google AI Studio, Google Antigravity, the Gemini enterprise platform and Android Studio.
• Researchers: academic and independent scientists can apply for the highest Antigravity usage tier free for 6 months. Eligibility, capacity and selection apply.
• Security teams: cyber defenders can apply to the Fairwind Program, which is expanding to "hundreds of additional" partners. The page implies the full cyber capabilities are gated behind this programme ("more trusted defenders").

Limits

• Nearly every number comes from a partner's internal benchmark. Most comparisons are against unnamed "peer models" or "every model we tested", so the multipliers (1.5×, ¼ cost, ~100×, 90%) cannot be checked or compared.
• The medical result uses synthetic data, and the partner says real-record validation is ongoing. The LexisNexis results are labelled "preliminary".
• The page has no first-party benchmark table, context-input size, latency figures or pricing. The "Performance" section is a heading with no figures.
• The safety measures (chain-of-thought and action monitoring that halts execution, refusal of harmful requests) mean some agent runs may be stopped or some dual-use requests refused. The page gives no detail on how often this happens.