Google opens Gemini 4 Argon to all; API price doubles December 1
The introductory rate matches mid-tier models from Anthropic and OpenAI. From December 1 it moves to $4 and $20, the same as their top models.
Conny Schneider / UnsplashGoogle made Gemini 4 Argon, its newest top-end model, available to every Google Cloud customer on October 13. Google first showed the model on September 30, with access limited to selected customers. Companies can call it through the Gemini API on Google Cloud (model ID gemini-4-argon), in Google AI Studio, in the Antigravity coding tool and in eligible Gemini Enterprise subscriptions.
The price is the part to plan around. Through November 30, Argon costs $2 for each million tokens sent in and $10 for each million tokens it writes, with cached input at 95% off. From December 1 the standard rate applies: $4 and $20. Any team that builds a business case on the November price will see its token bill double six weeks later.
How the price compares
VentureBeat lined up Argon's rates against current models from other labs:
| Model | Input, per 1M tokens | Output, per 1M tokens |
|---|---|---|
| Gemini 4 Argon, to November 30 | $2.00 | $10.00 |
| Claude Sonnet 5.5 | $2.00 | $10.00 |
| GPT-6.1 Sol | $2.00 | $10.00 |
| Gemini 4 Argon, from December 1 | $4.00 | $20.00 |
| Claude Opus 5.5 | $4.00 | $20.00 |
| Meta Muse Wonder | $4.00 | $20.00 |
| GPT-6 Astra | $10.00 | $50.00 |
So for the next seven weeks Argon is priced like a mid-tier model. After that it sits at the same rate as Anthropic's and Meta's top models, and well below OpenAI's GPT-6 Astra.
What the benchmarks say
Results are mixed, which is normal for a new model. On the independent Vals Index, which tests finance, legal, coding and tax tasks, Argon ranks first at 68.9%, just ahead of Claude Sonnet 5.5 and Claude Opus 5.5, VentureBeat reported. On the Artificial Analysis Intelligence Index it scores 53, behind Claude Opus 5.5 at 58. On Terminal-Bench 4.0, a test of command-line work, Opus 5.5 leads 66.4% to 57.4%.
Google's own numbers lean on security and agents. It says Argon is its hardest model yet to trick with indirect prompt injection, the attack where hidden instructions in a web page or document take over an agent: attack success fell to 0.7% across 15 adaptive attempts in the Gray Swan test. Argon blocks offensive tasks such as exploitation, but Google is giving a version without those limits to vetted defenders through what it calls the Fairwind Program.
Customers quoted by Google include GitLab, Harvey, Figma and ServiceNow. ServiceNow says Argon raised task success by roughly 20% in its own tests, with the largest gains in HR and customer service. ServiceNow uses it for its first-line service desk agent.
What Google did not say
Google's blog post gives no regional availability beyond a global endpoint. That matters for companies with data residency rules. It also does not list the context window. VentureBeat reports that Google now describes a limit of up to 1 million tokens for input and output combined, which does not match the 1 million output token ceiling Google described in September, and that Google has not explained the gap. Google says the model card and developer guide carry token limits and migration steps.
What IT teams should do
- Cost any Argon pilot at the December rate of $4 and $20 per million tokens, not the introductory one.
- Turn on context caching for repeated prompts. At 95% off cached input, it is the largest saving available.
- Before production use, confirm which region processes your data. The launch post lists only a global endpoint.
- If you already run Claude or GPT models, test Argon on your own tasks. Public benchmarks put the three within a few points of each other, and they disagree on which leads.
- If you build agents that read email, tickets or web pages, test Argon's prompt injection resistance against your own red-team prompts. Our AI strategy guide covers how to set the approval rules.