Google is back in the frontier AI conversation. On Wednesday, the company announced Gemini 4 Argon, its newest large language model, and claimed it outperforms OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1 across knowledge work, coding, science, math, and cybersecurity benchmarks. CEO Sundar Pichai said teams inside Google are already using it heavily, “from coding to quantum computing,” and promised to make it available “as soon as we can and as safely as we can.”
Investors liked what they heard. Alphabet shares rose about 1 percent in regular trading and added roughly another 1.5 percent in the extended session, according to MarketWatch.
What Google claims Argon can do
The benchmark claims come from Google’s own testing, so treat them the way you would any vendor scorecard. With that caveat, the company says Argon leads on the tasks that matter most to paying customers: knowledge work, AI-assisted coding, science and math reasoning, and cybersecurity.
The timing matters. After a stretch of delayed releases, the departure of prominent AI researchers, and poor benchmark showings, investors had started to wonder whether DeepMind was falling behind its rivals. Argon does not settle that question, but it is Google’s first strong counterargument in months.
One differentiator stands out on paper. Evercore ISI analyst Mark Mahaney notes Argon’s industry-leading output allowance of 1 million tokens. Long outputs are where complex agent workflows live, so a bigger ceiling at a lower price is a genuine selling point for anyone running multi-step tasks.
What it costs, and the price table that matters
Google launched Argon at an introductory $2 per million input tokens and $10 per million output tokens. Mahaney calls the pricing “highly competitive”: it undercuts GPT-6 Astra and Anthropic’s premium models while matching xAI’s Grok 4.7 on short-context input pricing.
Here is how the frontier stack looks on per-million-token list pricing:
| Model | Input | Output |
|---|---|---|
| Gemini 4 Argon (introductory) | $2 | $10 |
| GPT-6.1 Sol | $2 | $10 |
| Claude Sonnet 5.5 | $2 | $10 |
| Gemini 4 Argon (after intro) | $4 | $20 |
| Claude Opus 5.5 | $4 | $20 |
| GPT-6 Astra (standard) | $10 | $50 |
That table has a footnote that changes everything. Google says the $2/$10 rate is introductory and that pricing will double over time to land on par with Claude Opus 5.5, at $4/$20. Nobody has said how long the introductory period lasts.
The pattern is now the industry’s default playbook. OpenAI’s GPT-6.1 Sol and Anthropic’s Claude Sonnet 5.5 both launched at exactly $2/$10 a day apart last week, each promising near-flagship results without flagship bills. Argon joins that tier at launch, then graduates to the Opus 5.5 tier whenever Google decides.
The catch: almost nobody can use it yet
Argon is not launching as a broadly available developer model. Google is rolling it out first to trusted cyber defenders through its Fairwind Program, while the company participates in the U.S. government’s voluntary pre-release model access process. Broader access is planned for paid API customers and Google AI Ultra subscribers first, then developers, enterprises, and consumers.
That makes the benchmark lead theoretical for most buyers right now. Until developers can test Argon on real coding, legal, financial, and security workloads, nobody outside Google can verify whether the benchmark advantage turns into lower production costs.
What this means for the model race
Three things are worth taking away. First, the price war is structural, not promotional: every major lab now has a $2/$10 near-flagship tier, which tells you where the real production traffic lives. Second, Google’s benchmark lead is claimed, not independently verified, and its value depends entirely on when general access arrives. Third, the limited rollout to cyber defenders first is a deliberate trust play, positioning Argon as the responsible frontier model at a moment when Anthropic’s own IPO filing is warning investors about catastrophic AI risks.
Argon keeps Google in the race. Whether it wins anything depends on the one number Google has not given: the date.
Sources: VentureBeat, Barron’s, MarketWatch, reporting on Google’s September 30 announcement.
