ModelGoogle DeepMindConfirmed
Story 1 of 8 in the Tuesday 13 October editionGoogle DeepMind releases Gemini 4 Argon to developers and paid subscribers
The frontier model, first given to trusted cyber defenders on 1 October, is now on the paid Gemini API at an introductory price, and in Google Antigravity, the Gemini app and AI Mode in Search.

Google DeepMind made Gemini 4 Argon globally available on 13 October. Developers and enterprises can use it through the Gemini API in Google AI Studio and the Gemini platform. Google AI Ultra subscribers also get it in Google Antigravity, an agentic development platform, and AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search.
Google DeepMind says Argon ranks first on the Vals Index, which tests models on finance, coding, legal and tax work. Inside Google, teams using Argon in Antigravity raised their agents’ code submissions by more than 35% and cut rollbacks by nearly a third. The model reads text, images, audio, video and PDFs and replies in text.
Google DeepMind’s comparison with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 puts Argon highest or joint highest on 14 of 19 results.

Key numbers
- Introductory API price until 30 November 2026: $2 per million input tokens and $10 per million output, thinking included
- From 1 December 2026: $4 per million input tokens and $20 per million output
- Cached input: 95% cheaper, through implicit caching only
- Context window: 1,048,576 tokens for input, thinking and output combined, with replies beyond 64k tokens through long decode
Safety and access
The model card says Google DeepMind cannot rule out that Argon has reached the critical capability levels for biological and cyber risks in its Frontier Safety Framework, levels at which a model without safeguards could help cause severe harm, or a lower threshold for stealth and situational awareness. It says its safeguards bring the risk below each of them.
Cyber defenders in the Fairwind Program, which had Argon first, can apply for a version without cyber guardrails, and the programme is expanding to hundreds more partners. Google DeepMind is also offering 10,000 scientists the highest usage tier in Antigravity free for six months, subject to eligibility, capacity and selection.
In the Gemini API, the free tier and the Google AI Pro plan don’t include Argon. Google DeepMind says the rest of the Gemini 4 family will follow Argon, but hasn’t said when.
Sources
- PrimaryGemini 4 Argon: Built to solve your hardest problemsblog.google · Google DeepMind’s launch post, 13 October 2026
- OfficialGemini 4 Argon comes to the Gemini appblog.google · Availability for AI Pro and Ultra subscribers
- OfficialGemini 4 Argon model carddeepmind.google · Architecture and frontier safety results
- OfficialGemini 4 Argon on the Gemini APIai.google.dev · Model code, inputs and token limits
- OfficialWhat’s new in Gemini 4 Argonai.google.dev · API pricing, thinking levels and long decode
- OfficialGemini 4 Argon: our next era of frontier intelligenceblog.google · First announcement and benchmark table, 1 October 2026
- On X@GoogleDeepMind on X: Gemini 4 Argon is now globally availablex.com
- On X@sundarpichai (Sundar Pichai) on X: where Gemini 4 Argon is availablex.com
- On X@OfficialLoganK (Logan Kilpatrick) on X: rolling out Gemini 4 Argonx.com
- On X@GeminiApp on X: Gemini 4 Argon in the Gemini appx.com
- On X@GoogleAIStudio on X: Gemini 4 Argon in the Gemini API and AI Studiox.com
- On X@demishassabis (Demis Hassabis) on X: a wizard chess game built with Argonx.com
The rest of the day
Read the whole day02ProductApple
Apple unveils HomeView, a $299 home hub with its new Siri AI
HomeView has a 7-inch screen, recognises up to six people by face and runs Siri AI in beta.
03ProductGoogle DeepMind
Google DeepMind and Singapore build an on-device scam detector on Gemma
Announced on 14 October, the fine-tuned Gemma model reads messages on the phone and warns people when a conversation looks like a scam, without sending their chats to a server.
04Open sourceLiquid AI
Liquid AI releases LFM2.5-24B-A2B, an open model for agents on laptops and PCs
The reasoning model uses 2B of its 24B parameters for each token.
05InfraAnthropic
Anthropic replaces Git with Scion, its own source control built for agents
Daily merges at Anthropic rose more than tenfold between May and September 2026, so it built its own version control system and moved its monorepo off Git.
Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X







