Google DeepMind releases Gemini 4 Argon at $2/$10 per 1M tokens, claims #1 on Vals Index

Google DeepMind has released Gemini 4 Argon, its latest frontier model, with global availability starting today at an introductory price of $2 per 1M input tokens and $10 per 1M output tokens. Cached input tokens are priced at 95% off, which implies roughly $0.10 per 1M. Google says Argon ranks #1 on the Vals Index, a benchmark covering finance, coding, legal and tax work. That claim has not been independently verified.

Availability

According to Google, access is split by audience:

• Developers and enterprises: API in Google AI Studio and the Gemini platform
• Google AI Ultra subscribers: Google Antigravity
• Google AI Ultra and Pro subscribers: Gemini app and AI Mode in Search

Specifications

Item	Detail
Input price	$2 / 1M tokens (introductory)
Output price	$10 / 1M tokens (introductory)
Cached input	95% discount
Context window	Not disclosed in the announcement
Parameter count	Not disclosed
Training cutoff	Not disclosed
Numeric benchmark scores	Not published in the announcement; only the Vals Index #1 ranking is claimed

Google describes the price as "introductory" and does not say when it will change.

What Google says Argon does

Google positions Argon for multi-step work in coding, finance, legal, cybersecurity and science. It cites improvements in spatial reasoning, mathematical precision, visual coding, and multimodal document comprehension over earlier Gemini models. It also claims Argon follows style instructions better than any previous Gemini model, hedges less, and avoids repetition and clichés.

Internal results, as reported by Google

The following figures come from Google's own teams and have not been independently verified:

• Engineering throughput: Argon in Antigravity increased agentic code submissions by more than 35% and reduced code rollbacks by nearly a third.
• Chrome: Argon helped separate core webpage handling from a ~6 million line legacy folder inside Chromium's 40 million-line codebase. Google says this was a refactor across hundreds of files, completed in weeks rather than months.
• Agent Development Kit (ADK): Working autonomously overnight, Argon cut in-memory session prep time by 40% in long sessions and sped up cold imports by 21%. The changes are in google/adk-python.
• xTB quantum chemistry package: Up to 2.3x faster structure prediction and 1.8x faster conformer search, contributed upstream.
• Gemma 4 E2B: A 2.85x end-to-end inference speedup on an A100 via vLLM, from ~37,900 to over 108,000 tokens per second. Google reports 100% token parity and no quality drop.

Google also showed multi-agent demos in Antigravity that drive Blender via MCP, use Nano Banana for textures and Three.js for rendering, and build 3D scenes and games. One demo reconstructs 3D models of Chaco Canyon from roughly 50 GB of LiDAR, GIS, CAD and archival data.

What this means

The $2/$10 price is the most consequential detail. Pricing a frontier model at that level, with a 95% cached-input discount, makes long-running agentic workloads cheaper to run. Those workloads are the ones Google is emphasizing, such as overnight autonomous optimization and multi-agent loops. The word "introductory" is a caveat, and teams budgeting around it should plan for a change.

The evidence offered is thin on independent measurement. The only external benchmark claim is a #1 Vals Index ranking, with no score or margin published. The headline engineering results are Google's own, on its own codebases. Builders should wait for third-party evaluations, particularly on long-context behavior, since the context window was not disclosed.

The Gemma 4 E2B result is also notable. Google used Argon to speed up its own open model by 2.85x with token parity, which suggests the model can do low-level performance work such as attention tuning and speculative decoding, though that remains a vendor-reported result.