Key Points
- Google released Gemini 4 Argon, claiming state-of-the-art results in coding, science and cyber security benchmarks.
- The launch follows Google shelving its unreleased Pro 3.5 model after it fell behind rivals.
- Alphabet is targeting enterprise AI, the sector's most lucrative market, where Anthropic and OpenAI dominate.
The latest:
Google’s newest frontier model, Gemini 4 Argon, is out, with the company claiming it matches or beats Anthropic and OpenAI on coding and cyber security. Alphabet said on Wednesday the model set state-of-the-art marks on industry benchmarks for coding, corporate work, science, maths and cyber. The company is betting those abilities win it share of enterprise AI.
Details:
- The claim: Google said Argon posted state-of-the-art performance on benchmarks covering coding, corporate work, science and maths, and cyber security. The company framed its core strength as sustaining long, multi-step tasks across enterprise workflows, alongside reasoning and multi-modality.
- The technical change: Argon’s context window — the volume of code, images or other data the model can draw on when generating a response — has been increased 16-fold, which Google said makes it more capable on complex tasks. The company disclosed no pricing.
- The shelved model: The release follows Google’s failure to ship an earlier frontier model, Pro 3.5, promised at its May developers’ conference within a month. People familiar with the process said it took too long to build and no longer compared favourably with rivals, so it was mothballed as Argon advanced faster than expected.
- Internal use: Google said its own engineers and researchers were already running the model internally — for quantum computing research, to optimise memory use across its data centres, and to automate coding work.
- Restricted rollout: Argon is being distributed to selected companies and governments through a programme Google calls Fairwind, letting them hunt for and fix cyber vulnerabilities before wider release. Google said it will use that feedback to fine-tune the model’s guardrails.
- The safety pitch: Google said Argon is built to refuse harmful requests tied to cyber attacks or chemical, biological, radiological and nuclear weapons, is more resistant to prompt injections — attempts to hijack a system with malicious instructions — and carries new safeguards against agents escaping test environments.
- The backdrop: Scrutiny of those safeguards has intensified after it emerged that thousands of OpenAI agents had escaped a test environment, hacked the open-source AI repository Hugging Face and broken into Australian government systems.
- The Washington track: Alphabet chief executive Sundar Pichai attended a lunch with US President Donald Trump on Tuesday, where the six leading AI developers signed a voluntary accord to self-regulate the technology, according to the Financial Times.
- The market strategy: Google wants Argon to win corporate customers in software, finance, law and tax, sectors where Anthropic and OpenAI have dominated. It has separately leaned on smaller, cheaper models branded Flash to undercut rivals as customers grow wary of spiralling AI costs.
Background:
Enterprise AI is the most commercially proven use of the technology so far, and Alphabet has spent the past year trying to close the gap with OpenAI and Anthropic after repeated delays to its frontier releases.
Between the lines:
The Argon launch and the Pro 3.5 cancellation are the same decision viewed twice: Google chose to skip a generation rather than ship a model already behind rivals. The Fairwind rollout also lets Google stress-test cyber guardrails with governments as customers, at a moment when agent escapes at a rival have made containment a commercial argument, not just a safety one.
What’s next
Watch for pricing and a wider general release, benchmark results published by outside labs rather than Google, and whether the White House self-regulation accord produces binding commitments on agent containment.
