Google made Gemini 4 Argon generally available on October 13 at an introductory API price of $2 per million input tokens and $10 per million output tokens. Google’s own footnote sets the price at $4 and $20 once the introductory period ends on November 30, 2026. Any workload sized on the launch rate will cost twice as much a few weeks after it ships.
What Google released
According to Google’s announcement, Argon is available through the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Google Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. Enterprises can use it in Gemini for business, the API and Antigravity. Google says cached input tokens are priced at 95 percent off. It also reports a 1M unified input-output token limit through the Gemini API, with a new continuation protocol that developers need to set up.
Google says Argon ranks first on the Vals Index, a measure of model performance on finance, coding, legal and tax work. That ranking is Google’s claim and the post links to the benchmark rather than to an independent test of Argon in production.
The price has a date on it
Google’s launch post gives the introductory rate in its first paragraph and the standard rate in a footnote. Per that footnote, $4 per million input tokens and $20 per million output tokens apply from the end of November 30. Ars Technica reported the same schedule and noted that developers will have to factor it in before building workflows around the model.
Our arithmetic shows how fast that compounds. A team that sends 1 billion input tokens and receives 200 million output tokens in a month pays $4,000 at the introductory rate ($2,000 for input, $2,000 for output) and $8,000 at the standard rate. Cached input is billed at a fraction of the input price, so a workload that reuses long context pays less on both schedules. The ratio between the two schedules is still two to one on every uncached token.
A pilot that runs from mid-October to the end of November will therefore measure a price that no production contract will see. Finance teams can handle this by modeling the standard rate from the first day and treating the discount as a one-time credit. Procurement can ask whether committed-use terms exist for Argon and whether the standard rate is fixed for any period after launch. Google’s post answers neither question.
Access depends on the plan tier
The same release splits access by subscription. Ars Technica reports that the Gemini app offers Argon to AI Pro ($20 a month) and AI Ultra (from $99 a month) subscribers, and that in Antigravity and AI Studio the model is limited to AI Ultra plans. The report adds that Google has declined to say when or whether Argon will reach Antigravity on other tiers. Google’s post lists the same destinations for each tier without stating a timeline for the rest.
The Decoder reports that free Gemini users and Plus subscribers have access only to Flash models under a new tier structure, and that Argon replaces the Gemini Pro class as Google’s frontier series. Ars Technica says Google is keeping Gemini 3.1 Pro available for the time being, which matters to any team with a prompt library tuned on it. Ars Technica quotes no retirement date, and the Google post gives none.
For a buying committee, this means the model engineers can reach on a team plan and the model the API serves under contract are different questions. Developers who prototype in a subscription product may be testing a tier their company does not license for production.
The controls Google documents
Google devotes a section to safety and security, and several of its claims bear on agent deployments. It says Argon was trained to respect system boundaries, report negative findings honestly and pause for human confirmation before taking irreversible steps. On coding benchmarks without an incentive to hold back, Google reports shortcut attempts fell to 2 percent. It also describes an automated misalignment monitor that reads an agent’s reasoning and actions and blocks unintended workarounds in real time.
On prompt injection, Google says Argon leads the Gray Swan Indirect Prompt Injection benchmark, cutting attack success across 15 adaptive attempts to 0.7 percent. These figures come from Google’s own testing. A committee should ask for the model card Google references and decide which tests to repeat on its own data before relying on them.
One program deserves a policy decision. Google says defenders can apply to the Fairwind Program for access to Argon without cyber guardrails, and that it is expanding the program to hundreds of additional partners. Security leaders whose vendors or contractors take part should know what that access covers and how it is audited. Our earlier coverage of Google’s Gemini agent with spend caps and agent identities described the governance controls Google ships around its agents.
What it means for the buying committee
Frontier model pricing now carries an end date. Our coverage of Sonnet 5.5 and token efficiency argued that tokens consumed per completed task decide what a model costs, whatever the list price says. Argon adds a second variable, which is the date on which the list price changes.
Three actions follow for a CIO or chief AI officer evaluating Argon.
- Budget every Argon workload at $4 and $20 per million tokens, and treat any saving before December 1 as a credit.
- Confirm which subscription tier and which API tier each team is actually testing on, and write that into the pilot record.
- Ask Google in writing for committed-use terms, the Gemini 3.1 Pro retirement date, and the scope of the Fairwind Program, none of which the launch post states.
Google’s own launch figures describe teams finishing months of engineering work in weeks. Those are Google-reported results on Google’s codebases, so a committee should read them as a prompt to run a bounded pilot against its own work with a standard-rate budget attached.
Source: Google Blog
