Google finally has a new flagship AI model. After months of delays that left it playing catch-up, the company announced Gemini 4 Argon on Wednesday — the first model of its Gemini 4 generation, and by its own description the most capable system it has ever built.
Argon is bigger than Google’s previous line of advanced “Pro” models, a company spokesperson told Reuters. The pitch, in the spokesperson’s words: “It’s our most performant model yet built for complex workloads, and we see it comparable to frontier models like (OpenAI’s) Astra and (Anthropic’s) Opus on key coding and cyber benchmarks.” The launch price is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off — aggressive pricing for a frontier-tier model, and a clear signal Google wants volume once it opens the gates.
The gates, however, are staying mostly shut for now.
Google’s comeback attempt: what Gemini 4 Argon actually is
Argon is built to sustain deep reasoning across complex, long-horizon workflows, Google said in a Wednesday blog post — tasks that stretch over hours rather than a single prompt-response cycle. That includes real-world software engineering, enterprise knowledge work like legal and finance research, and cybersecurity defense. The company says its own engineers are already using it day to day for debugging and large codebase migrations, and that it can parse visuals too, from long videos to dense charts.
But you can’t use it yet. Not even close.
Google is releasing Argon only to a select group of cybersecurity partners through its Fairwind Program, the company’s security initiative, while it also takes part in the Trump administration’s voluntary process for pre-release model access. “Safely releasing frontier capabilities at this level requires a phased approach,” the Google blog said, adding that the company will expand access “to developers, enterprises, and consumers as soon as possible” — without saying when. No timeline for a public release was provided at all.
That’s a deliberate choice, and a defensive one. For a model whose flagship skill is defensive cyber work (Google says Argon can “autonomously find, validate, and patch critical software vulnerabilities”), the offensive misuse potential is obvious. Giving it first to trusted defenders and to Washington, before general availability, reads like an attempt to set the safety narrative on Google’s terms this time.
A cyber specialist first, everything else later
The cybersecurity-first positioning is the most interesting thing about this launch. Every lab markets its newest model as the smartest; Google is marketing Argon as the safest and the most useful to the people who patch software for a living. It claims Argon scored significantly higher than OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus models across a range of benchmarks, and points to Vals, an increasingly popular benchmarking startup, whose AI model index currently ranks Argon first.
Take the benchmarks with the usual grain of salt. These are self-reported numbers, and Google’s own release shows Argon trailing on other metrics — including two of the four coding-related benchmarks it chose to include. Vendors publish the tests they win. Still, even placing third or fourth in coding while leading in defensive cyber work would make Argon a distinctive product rather than a me-too frontier model. Nobody else is launching their flagship as a vulnerability-hunter first.
CEO Sundar Pichai took to X to frame the rollout as the responsible kind: “Importantly Argon has frontier safeguards and we are rolling it out responsibly — it’s with the US gov’t and going to a set of trusted cyber defenders through our Fairwind Program today.” He wants the story to be about caution as much as capability.
What Google killed to ship this
Argon arrives trailing a long shadow. Google confirmed it no longer plans to release Gemini 3.5 Pro, the model Pichai originally said would arrive in June. It just never shipped. In the gap, Google overhauled its DeepMind AI lab: founder and CEO Demis Hassabis stepped aside, and several leaders of the Gemini effort left the company, according to Reuters.
That’s an extraordinary amount of churn for one product cycle, and it explains the delay better than any technical excuse. The model’s training didn’t fail; the team building it got rebuilt mid-flight. Meanwhile Anthropic and OpenAI kept shipping: Anthropic with Fable earlier this year and Opus, OpenAI with Astra, which the company hailed as its best model yet. Google watched both labs pull ahead in the public’s mind while it had nothing new to show. Argon is the answer to that. It’s late, it’s cautious, and it’s narrow in distribution. But it’s finally something.
Why this matters
The honest read: Google is trying to change what the race is about. It can’t win the “who shipped first” contest anymore. It lost that, badly. So it’s trying to win the “who shipped the model that actually helps your security team” contest instead. Defensive cybersecurity is a domain where deep reasoning across long workflows genuinely matters, where the buyer is an enterprise CISO rather than a developer on a free tier, and where Google’s cloud business gives it distribution the other labs don’t have.
Whether it works depends on two things Google hasn’t proved yet: that Argon is meaningfully better at cyber defense than Astra or Opus (self-reported benchmarks aren’t proof), and that the company can ship broadly without the next round of internal upheaval. The talent story cuts both ways — new leadership can unstick a stalled lab, but it can also mean the roadmap gets rewritten again in six months.
For now, Argon is a promise made to a very small room: cyber defenders, the US government, and Google’s own engineers. The rest of us get the blog post and the price list. Whether that room gets bigger, and how fast, is the story to watch.
Frequently asked questions
What is Google Gemini 4 Argon? Gemini 4 Argon is Google’s new top-tier AI model and the first release of its Gemini 4 generation. It’s built for long, complex tasks in cybersecurity defense, software engineering, and enterprise work like finance and legal research, and Google says it’s larger than its previous Pro-class models.
When will Google Gemini 4 Argon be available to the public? Google hasn’t announced a public release date. Argon is initially rolling out only to select cybersecurity partners through Google’s Fairwind Program and to the US government through a voluntary pre-release access process, with wider access to developers, enterprises, and consumers promised later.
How does Gemini 4 Argon compare to OpenAI’s Astra and Anthropic’s Opus? Google claims Argon beats Astra and Opus on several self-reported benchmarks, including coding and cyber tests, and cites the Vals benchmark index showing Argon on top. But it also lagged its rivals on other metrics, including two of the four coding benchmarks Google included in its own release.
How much does Gemini 4 Argon cost to use? Google announced an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted 95% off the input price.
Sources: Reuters, TechCrunch, Google