Google launches Gemini 4 Argon amid access and benchmark disputes

Google released paid Gemini 4 Argon on Oct. 13, 2026, at $2/$10 intro pricing, with disputed access and mixed independent benchmarks.

Google and Google DeepMind said they released Gemini 4 Argon worldwide on October 13, 2026, as a sparse mixture-of-experts frontier model trained on TPUs, with no open weights, a combined 1 million-token window, and maximum output of up to 1 million tokens, up from 64K. Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off, then $4 and $20 after November 30, 2026. Sources disagree on access: Google says Pro and Ultra subscribers get the Gemini app and Search AI Mode and Ultra subscribers get Antigravity, while Ars Technica says Antigravity and AI Studio are Ultra-only, the app is limited to AI Pro at $20 a month and AI Ultra from $99, and free users stay on Flash-Lite; The Decoder says free and Plus users receive only Flash models, and Ars says Gemini 3.1 Pro remains available. Google says Argon leads the Vals Index and cites internal gains of more than 35% more agentic code submissions, nearly a third fewer rollbacks, and a 2.85x Gemma 4 E2B speedup past 108,000 tokens per second, but other reports say it trails Claude Opus 5.5 overall, matches that model at 52.8% on CursorBench, sits near GPT-6 Astra at about 60% of the cost, and finishes last of four at 55.0% on FrontierSWE v2 despite 91.7% on LVBench. DeepMind says the broader rollout follows Fairwind access for defenders, that unguarded testing solved all 12 simulated cyber scenarios versus five earlier, and that new screening and monitoring cut safety violations nearly twofold, alongside a 0.7% Gray Swan result, under 1% of tested bio sub-tasks unlocked, and a model card that cannot rule out critical biological or cybersecurity capability levels.

• Google and DeepMind made Gemini 4 Argon available worldwide on October 13, 2026, via the Gemini API, AI Studio, the Gemini app, Search AI Mode, and Antigravity.
• Argon is a sparse mixture-of-experts multimodal model trained on TPUs, with no open weights or disclosed parameter count, a combined 1 million-token window, and maximum output raised from 64K to as much as 1 million tokens.
• Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% off, rising to $4 and $20 after November 30, 2026.
• Access is disputed: Google says Ultra and Pro subscribers get the app and Search and Ultra gets Antigravity; Ars Technica limits the app to AI Pro ($20/month) and AI Ultra (from $99/month), limits Antigravity and AI Studio to Ultra, and…
• Google says Argon ranks first on the Vals Index; The Decoder says it trails Claude Opus 5.5 in independent overall rankings. Other scores include 91.7% on LVBench, 77.9% on DeepSWE v1.1, 55.0% on FrontierSWE v2 (last of four), 52.8% on…
• Google cites agentic code submissions up more than 35%, rollbacks cut nearly a third, 40% faster ADK session prep, up to 2.3x faster xTB predictions, and Gemma 4 E2B inference 2.85x on an A100 via vLLM, from about 37,900 to over 108,000…
• DeepMind expanded Fairwind, including Wiz; unguarded tests solved 12 of 12 simulated cyberattack scenarios versus five before. Safeguards cut adversarial safety violations nearly twofold; Gray Swan indirect prompt injection succeeded on…