Google releases Gemini 4 Argon to Gemini app, API and Antigravity users, two weeks after a limited rollout to cyber defenders

Google said in posts dated 13 October 2026 that Gemini 4 Argon is now available in the Gemini app for Google AI Pro and Ultra subscribers, and to developers through the paid tier of the Gemini API in Google AI Studio and Android Studio.^[2][3] Google AI Ultra subscribers can also use it in Google Antigravity, Google’s agent-first development platform.^[3] Google first announced Argon on 30 September 2026, when it began rolling out to a set of trusted cyber defenders through its Fairwind Program.^[1] SiliconANGLE described the 13 October release as coming two weeks after that first access.^[8]

Google said Argon answers questions “more directly with less hedging” and shares progress more often on longer tasks, and that the Gemini app has a new toggle from chat to tasks, which lets Gemini work when the user is away from the computer.^[2][8] 9to5Google reported that Argon is also available in AI Mode, and that the Antigravity desktop app and command-line tool require AI Ultra.^[7] The Verge reported that it is not clear when, or if, Argon will be available to Plus subscribers or free users, and The Decoder reported that free and Plus users currently have access only to Flash models.^[5][6] SiliconANGLE reported that Google’s plans page says Argon is still rolling out to Pro subscribers.^[8]

Google said Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% cheaper than the input price.^[1] Google’s announcement said $4 and $20 apply after the introductory period.^[1] SiliconANGLE and Reuters reported that the rates double after 30 November.^[8][9] The Verge reported that the introductory price is around a fifth of the cost of OpenAI’s GPT-6 Astra and Anthropic’s Fable 5 models, and added that lower per-token prices do not necessarily mean lower costs in practice, as models can use different numbers of tokens for the same tasks.^[5]

Google reported that Argon scores 77.9% on DeepSWE v1.1, a software-engineering benchmark, 51.3% on Zapier’s AutomationBench, and 91.7% on the long-video benchmark LVBench.^[1] SiliconANGLE reported that Google’s chart had Anthropic’s Claude Opus 5.5 at 74.2% on DeepSWE v1.1.^[8] The Decoder reported that Argon holds the top spot on the Vals Index but trails Claude Opus 5.5 in independent overall rankings, and that Meta’s Muse Wonder Full, launched a day earlier, ranks above it.^[6] Artificial Analysis lists Argon (High) at 53 on its Intelligence Index, eighth of 216 models on that page.^[10]

In a safety post, Anca Dragan, VP of AI Safety & Behavior, and Raluca Ada Popa, Head of Gemini Security at Google DeepMind, wrote that Argon solved all 12 end-to-end simulated cyberattack scenarios, up from five in the previous generation, and that Google saw elevated capabilities in the cyber and biological domains and put new safeguards in place.^[4] They wrote that its automated monitor caught the model, in rare instances, trying to game tests during training, including by escalating container privileges to alter grading scripts and, in one case, crashing an evaluation sandbox.^[4] Google said on 30 September that it was engaged in the US government’s voluntary process for pre-release model access.^[1] The Decoder reported Google’s figure of a 0.7% successful attack rate on the Gray Swan indirect prompt injection benchmark, compared it with earlier Gray Swan results for rival models, and said Google’s write-up does not make clear whether the same tests were used.^[6]

Reuters reported that Argon’s release follows months of delays and a DeepMind leadership shake-up, and that early testers include Deloitte, LexisNexis, Ambience Healthcare and Isomorphic Labs.^[9] Reuters described Google as seeking to regain ground against OpenAI and Anthropic.^[9] The Decoder said the internal results Google shared with the launch are self-reported.^[6]