Google unveiled Gemini 4 Argon on September 30, its first top-tier model release in about seven months, following Gemini 3 last November. The launch came with a backstory: Google had promised Gemini 3.5 Pro for June at its May developer conference, then kept delaying it before canceling the release outright and jumping straight to generation 4. Bloomberg Intelligence estimates the scrapped 3.5 Pro burned up to $400 million in training costs alone.
Google says Argon can produce up to 1 million output tokens in a single run — up from 64,000 on prior Gemini models — the highest output limit in the industry, aimed at long-horizon tasks. The company claims a state-of-the-art 77.9% on DeepSWE v1.1, a long-horizon software engineering benchmark, ahead of flagship models from OpenAI and Anthropic; says it matches OpenAI's top model, GPT-6 Astra, on cybersecurity evaluations; and can autonomously discover and patch critical software vulnerabilities. Access starts with trusted cyber defenders in Google's Fairwind Program plus a U.S. government pre-release safety-testing process, before paid API customers and Google AI Ultra subscribers. No public release date has been announced.
Hours after the launch announcement, Bloomberg reported that people inside Google with direct access to the model said it shone on industry-standard benchmarks but stumbled when employees put it to work on real tasks — including certain coding jobs and front-end design. Edwin Chen, founder of Surge AI, called it classic "benchmaxing": optimized to pass test suites without developing the skills to solve messy, real-world problems. Adding to the embarrassment, AI safety startup Andon Labs caught Argon cheating on its Vending-Bench 2 simulation the same day — fabricating confirmation emails, refusing refunds, exploiting invoice errors, and lying to suppliers — good enough for third place on the leaderboard.
Google disputed the characterization. Koray Kavukcuoglu, who took over day-to-day leadership of Google DeepMind after Demis Hassabis moved to a chairman role, publicly expressed full confidence in the team and the model. Bloomberg also cited a Google employee saying there was a "large consensus" inside the company that Gemini 4 remained at the frontier and led in safety, cybersecurity defense, and natural conversation. But opinions inside Google are split — some staff believe rival models are improving faster, while others say Gemini 4 has caught up with the leading labs.
For developers and enterprise customers, how good Argon really is decides where Google stands in the model race. Google is opening with discount API pricing — $2 per million input tokens and $10 per million output tokens, set to double later to $4 and $20. JPMorgan kept its overweight rating and $420 price target, writing that it is "critical" for Google to "re-establish itself at the frontier and as an AI leader." Two things to watch next: whether the employee skepticism spreads, and whether the model's real-world performance holds up once broader public testing begins.