Gemini 4 Argon is rolling out first to cybersecurity defenders, with company-reported benchmark leads and an introductory price of $2 per million input tokens. By early afternoon, Alphabet was down about 1.4%, roughly $60 billion of market value.
Alphabet's investors got their first full session with Gemini 4 Argon on Thursday. They started buying and finished selling.
Class A shares opened at $350.79, up 1.95% from Wednesday's close of $344.08, and reached $353.22. By the 12:30 p.m. half-hour they had fallen to $335.51. At about 2 p.m. they traded near $339.13, down 1.44%. With about 12.1 billion shares outstanding, the decline is worth roughly $60 billion, and the stock stood about 4% below its morning high.
What Google launched
Google introduced Gemini 4 Argon on Wednesday and described it as its new frontier model. The first users are "a set of trusted cyber defenders" enrolled in a program called Fairwind. Those users receive it "without cyber guardrails." A broader release to developers, enterprises and consumers will follow, starting with paid API customers and Google AI Ultra subscribers, on a timetable Google has left open.
The company said Argon "sets a new state of the art" on the DeepSWE v1.1 coding benchmark, with a 77.9% score, and leads the Vals Index and AutomationBench. Its output limit expands to 1 million tokens from 64,000.
The price is the economic headline. Argon launches at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after an introductory period. Analysts at Jefferies put Anthropic's Claude Opus 5.5 at $4 and $20 and its Claude Fable 5.1 at $10 and $50, which leaves Argon's introductory tier at half of Opus's list price.
The doubts
Some Google employees with direct access to the model say it performs better on benchmarks than in practical use and struggles with certain coding tasks, according to people familiar with the matter. Google also dropped a planned June release called Gemini 3.5 Pro, those people said. Google said it would be "inaccurate to say that Gemini 4 is underperforming in areas such as coding."
Jefferies noted that Argon trails leading competitors on the more demanding FrontierSWE v2 benchmark and tied with Grok 4.7 at 68% on CWE-bench v1. "Starting this rollout in this way gives us more confidence," said Tulsee Doshi, a product lead for Gemini models.
A crowded two days
Alphabet's sharpest drop on Wednesday came in the final half-hour, when the stock fell 1.28% on 7.1 million shares, against 0.6 million to 1.2 million in earlier afternoon bars. That bar followed word of employee skepticism about the model and confirmation that the Federal Trade Commission had opened an inquiry into OpenAI, Anthropic and other AI developers. Google was not named in the FTC inquiry. The stock then rose in after-hours trading once the launch was public.
Thursday's reversal coincided with a 10-year Treasury yield at its highest since 2002 and a rotation into software stocks after Accenture's results. Peers were mixed: Microsoft rose 0.45%, Meta 0.43% and Nvidia 0.85%, while Amazon slipped 0.08% and Broadcom fell 1.21%. Alphabet lagged the Nasdaq 100 tracker by about 1.6 points.
Analysts
JPMorgan, which rates the stock Overweight, said Argon "shows frontier performance in complex workflows, cyber defense, and software engineering," and that Google needed to "re-establish itself at the frontier." KeyBanc, also Overweight with a $435 target, called the launch earlier than expected but said the limited rollout makes real-world performance hard to judge. Jefferies rates the shares Buy with a $445 target. Morningstar, with a $433 fair value estimate, said its valuation already assumes a return to the frontier.
Two readings
One reading is that Argon puts Google back among the leading models and that an aggressive price can win developer and enterprise spending, supporting Google Cloud and Gemini products across Search, Workspace and Android.
A second reading is that a gated launch resting on company benchmarks leaves the competitive question open, and that cutting price to compete points to thinner margins at the model layer.
The test
General availability is the event that allows outside evaluation. Until paid API customers can run Argon on their own work, independent leaderboard results will be thin. Alphabet's third-quarter report, and whether usage and pricing figures appear in it, comes after that.