Artificial Intelligence
Google launched Gemini 4 Argon
Google said Gemini 4 Argon became available across developer, enterprise and consumer products, while claiming it was already improving software engineering, security, science and robotics work.
- Google said Gemini 4 Argon was available through API access in AI Studio and the Gemini platform, and to subscribers in Google Antigravity, the Gemini app and AI Mode in Search.
- Google said the model started at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.
- Google said Argon ranked first on the Vals Index and outperformed other frontier models on benchmarks across several sectors.
- Google said internal use of the model raised agentic code submissions by more than 35% and reduced code rollbacks by nearly a third.
- Google also said partners and researchers used Argon in work for Wiz, Gamow Labs, Apptronik and Professor Matt Landreman at the University of Maryland.
Google said Gemini 4 Argon was available worldwide through the API in AI Studio and the Gemini platform, and for Google AI Ultra subscribers in Google Antigravity. It also said Google AI Ultra and Pro subscribers could use it in the Gemini app and in AI Mode in Search.
The company said the model was aimed at coding, finance, legal work, cybersecurity and science. It also said it placed first on the Vals Index and that its own teams had seen more than a 35% rise in agentic code submissions, alongside a fall of nearly a third in rollbacks.
In examples from Google, Argon helped Chrome separate core webpage handling from a legacy browser-features folder in Chromium. Google also said it improved work on its open-source Agent Development Kit, sped up xTB and Gemma 4 E2B, and supported 3D visualisation work.
Google said Wiz used Argon in its Scan for Good initiative to find and fix a data-exposure gap, and that the same work involved a federal agency. It also said the model helped Gamow Labs analyse de-identified newborn genomes, and that Google and Professor Matt Landreman of the University of Maryland found broader counterexamples to Grad’s conjecture in fusion research.
In robotics, Google said Argon watched a video of a person playing cornhole, built simulation code and was then transferred to Apptronik’s Apollo humanoid, which made seven consecutive shots in a trial. The post ended by saying Google would widen access through Fairwind Program partners and offer 10,000 scientists six months of free top-tier use in Antigravity, subject to eligibility and capacity.
Named in this story
People
- Professor Matt Landreman
- co-discovered broader counterexamples to Grad’s conjecture with Google
Companies
- published the launch post and attributed the model claims to itself
- Wiz
- used Argon in its Scan for Good initiative and found security issues
- Gamow Labs
- used Argon to analyse de-identified newborn genomes
- Apptronik
- the company behind the Apollo humanoid used in the robotics trial
Organisations
- Fairwind Program
- the programme Google said it was expanding to more partners
- University of Maryland
- Professor Matt Landreman’s affiliation
Products and systems
- Gemini 4 Argon
- the model that was launched worldwide
- Google AI Ultra
- the subscription tier that unlocked access in some Google products
- Google Antigravity
- one of the products where the model was made available
- Gemini app
- one of the apps where subscribers could use the model
- AI Mode in Search
- another Google surface where subscribers could use the model
- AI Studio
- one of the developer surfaces offering access to the model
- Gemini platform
- another platform offering access to the model
- Chrome
- used Argon in a browser refactor example
- Chromium
- the codebase that contained the browser features Google described
- Google’s open-source Agent Development Kit (ADK)
- the framework Google said Argon helped optimise
- xTB
- the quantum chemistry package Google said it sped up
- Gemma 4 E2B
- the model used in Google’s inference speedup example
- vLLM
- the serving stack in Google’s speedup claim
- AlphaQubit
- the quantum error-correction decoder Google said it translated into a hardware kernel
- Lean
- the proof assistant used in Google’s quantum verification example
- Apptronik’s Apollo humanoid
- the robot that made seven consecutive shots in the trial
How the source tells it
The article was a promotional product launch written in a self-congratulatory register.
- vendor boosterism first-person company claims and internal success stories were used to sell the model
- hype superlative benchmark claims and broad impact statements were repeated without independent support
- triumph a series of internal and partner examples framed the launch as a sequence of wins