DeepMind released Gemini 4 Argon, a next‑generation model Google says delivers
frontier performance on complex tasks including software engineering, enterprise
knowledge work (legal and finance) and cybersecurity defense. Initial access is
limited via the Fairwind program to trusted cybersecurity defenders and will
expand to developers, enterprises and consumers after additional safety
measures. Pricing is $2 per mln input tokens and $10 per mln output tokens;
cached input tokens are priced 95% below standard input tokens (implying about
$0.10 per mln). Argon supports up to 1 million‑token outputs, versus the prior
~64k token limit. Google reports internal uses for code development, research
and engineering optimization, including helping quantum computing teams tune
algorithms and analyzing data‑center operations to free more than 300 TiB of
memory. Google cites benchmark results of 77.9% on DeepSWE v1.1 software
engineering tests and 68% on CWE‑bench v1 vulnerability repair (tied for top
models). Before wide release Google says it will strengthen anti‑abuse systems,
prompt‑injection defenses, model behavior monitoring and sandboxing; Argon will
be made available to paid API customers and Google AI Ultra users.
