Google unveils Gemini 4 Argon: a frontier model shipped first to trusted cyber defenders, with a 1M-token output limit

Specs and price are unusually concrete this time. Google's blog states that Argon raises the output token limit from 64K to 1 million, which the company calls industry-leading, reasoning that headroom to generate hundreds of thousands of tokens in a single trajectory adds a new level of depth to reasoning. Pricing is equally explicit: an introductory $2 per million input tokens and $10 per million output tokens, with cached input priced at 95% off the input price; after the introductory period, $4 and $20 apply. The rollout order is stated too: cyber defenders and trusted testers first, then paid API customers and Google AI Ultra subscribers.

On capability numbers, Google lists several firsts: a new state of the art of 77.9% on DeepSWE v1.1, which measures real-world long-horizon software engineering; leadership on the Vals Index, which weights finance, coding, legal and tax work by contribution to US GDP; first place at 51.3% on AutomationBench, Zapier's benchmark for end-to-end execution across core business functions; and 91.7% on LVBench for long video understanding. Google also gives three internal use cases: in quantum algorithmic optimization, Argon beat a published baseline by 40% in minutes; a team of Argon agents analyzed fleet-wide profiling telemetry and autonomously applied memory optimizations, freeing over 300 TiB of memory once rolled out, with total savings estimated at 500 TiB to 1 PiB; and large-scale C and C++ to Rust migrations ranging from tens of thousands of lines in core libraries such as re2 and libgav1 up to 800K+ lines for the Fuchsia Zircon kernel.

Cybersecurity is both the entry point for this launch and the place where the edges matter most. Google says Argon can autonomously find, validate and patch critical software vulnerabilities, and that it will release a version without cyber guardrails to trusted defenders and its own internal teams so they can use the full capability. One disclosed early case: through Wiz's Scan for Good initiative, the model uncovered a critical vulnerability exposing sensitive personal information across healthcare software used by hospitals worldwide, which Google says previous frontier models had missed. On CWE-bench v1, which evaluates vulnerability remediation, Argon ties for first place at 68%. The other side to keep in mind: consumer-facing guardrails and refusal policies are not final, and Google lists that work as a prerequisite for a broader rollout.

On safeguards, Google lists four lines. First, defending against misuse: refusal policies against cyber and CBRN attacks, with robustness improved by monitoring internal activations, tested by internal and external red teams. Second, defending against indirect prompt injection, where Google calls Argon its most resilient model yet and says it leads on Gray Swan's IPI benchmark. Third, monitoring for misalignment: watching chain-of-thought and actions and stopping execution when necessary, while urging the industry to preserve reasoning transparency. Fourth, hardening systems: isolating and sealing sandboxed environments before high-risk training or evaluations, and sharing agent security practices with partners. All four are process commitments whose real value depends on the evaluations and incident disclosures that follow.

Split the competitive picture into what Google says and what reporters add. Google's own framing is that Argon is comparable to OpenAI's Astra and Anthropic's Opus on key coding and cyber benchmarks. Reuters adds three points: Argon is larger than the previous Pro line; the company gave no timeline for a public release and no longer plans to ship Gemini 3.5 Pro, which had been expected in June; and on the four coding-related benchmarks Google published, Argon trails on two, with all of these being company-reported results. For developers, the most practical takeaway is not who is first but the sequence for access: trusted testers, then paid API, then subscription tiers.