Google releases most advanced Gemini AI model

[Chip (OP)]

https://www.ft.com/content/46194a0b-a0e4-42cc-ad40-0df753492768?syn-25a6b1a6=1

Google releases most advanced Gemini AI model

Tech giant hopes ‘Argon’ will help it re-establish itself at the frontier of artificial intelligence

1st Oct 2026

Google has released a new version of its most advanced AI model that it claims matches or beats rivals Anthropic and OpenAI in coding and cyber security, as the tech giant attempts to recover its position as a leader in the sector.

The Silicon Valley group on Wednesday claimed Gemini 4 “Argon” had state of the art performance on industry benchmarks for coding, corporate work, science and maths, and cyber security. Parent company Alphabet is counting on these abilities to galvanise its market share in enterprise AI, the most lucrative use case for the technology.

Argon has been highly anticipated since Google failed to release an earlier version of its frontier model, 3.5 Pro, after promising at its annual developers’ conference in May that it would arrive within a month.

Google ultimately decided to skip 3.5 Pro because its training, which took longer than expected, produced disappointing results and its capabilities no longer compared favourably with rivals, people familiar with the process said. When Argon developed faster than expected over the summer, the decision was made to mothball its predecessor.

The company said its engineers and researchers were using the model internally for quantum computing research, to optimise its memory use in data centres and to automate coding.

Argon is more adept at complex tasks because of a far larger context window — the amount of data such as code or images that it can refer to when generating a response — which has been increased 16-fold.

“Gemini 4 Argon’s capabilities across coding, reasoning and multi-modality and its ability to sustain long, multi-step tasks enable it to excel across a range of enterprise workflows,” the company said.

Google aims to advertise these capabilities to help it capture a larger share of the lucrative corporate AI market in fields such as software, finance, law and tax. Anthropic and OpenAI have previously dominated in these sectors.

Google said the model was being distributed to select companies and governments through its “Fairwind” programme, which will allow them to use the model to identify and fix cyber vulnerabilities ahead of its wider release.

Google will use their feedback to fine-tune its guardrails and said Argon was designed to refuse harmful requests related to cyber or chemical, biological, radiological and nuclear attacks.

The company said it is more resistant to prompt injections — attempts to hijack a model using malicious instructions — and has added safeguards to prevent its agents escaping test environments.

Debate about such safeguards has accelerated in recent weeks after it emerged that thousands of OpenAI agents had conspired to escape a test environment and hack open-source AI repository Hugging Face, as well as break into Australian government systems.

Alphabet’s chief executive Sundar Pichai attended a lunch with US President Donald Trump on Tuesday, signing a voluntary “accord” between the six leading AI developers to “self-regulate” the technology.

“Argon has frontier safeguards and we are rolling it out responsibly — it’s with the US government and going to a set of trusted cyber defenders,” Pichai said. “We’re going to make it available as soon as we can and as safely as we can.”




The Black Market for AI Chips

https://forum.drugs-and-users.org/gemini.mp4