TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google introduced Gemini 4 Argon, its new frontier model, with early independent tests placing it alongside OpenAI models on one overall index and ahead of rivals on an agent benchmark. Anthropic’s models remain ahead on the Intelligence Index, while Argon’s lower introductory token prices are offset by higher output-token use on tested tasks. The model is initially restricted to selected cyber defenders and Google teams; public access has no announced date.
Google has introduced Gemini 4 Argon, a new frontier AI model whose early independent benchmark results put it level with some OpenAI and Anthropic systems on selected measures, but do not show a clear overall lead. The model is initially available only to a limited group of cyber defenders and Google teams, while the company has not given a date for broader access.
In assessments by Artificial Analysis, Argon scored 53 on the organization’s Intelligence Index at its High reasoning setting, tying GPT-6 Astra and Claude Fable 5.1. Anthropic’s Claude Opus 5.5 scored 58 and Claude Sonnet 5.5 scored 56. Argon’s result was 23 points above Google’s Gemini 3.1 Pro Preview, according to the same assessment.
Results varied by task. Argon ranked first on Artificial Analysis’s version of AutomationBench-AA, scoring 77.5%, six percentage points ahead of Claude Sonnet 5.5 at its maximum setting. On Terminal Bench 4, Argon scored 57%, substantially above Gemini 3.1 Pro Preview’s result but below Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra, which scored 64%, 60% and 59%, respectively.
Google’s initial access plan starts with selected “trusted cyber defenders” through its Fairwind program and the company’s internal teams. Google says feedback from early testing will inform safety measures before it offers the model to developers, businesses and consumers. The company has set an introductory API price of $2 per million input tokens and $10 per million output tokens, but says the wider release will begin with paying API customers and Google AI Ultra subscribers.
Benchmark Gains, With Trade-Offs
Argon’s results suggest Google has returned to the leading group on some demanding AI tasks after more than seven months without a new frontier model, according to The Decoder’s report. Its strong score on AutomationBench-AA may matter to users evaluating models for agentic work, where systems carry out multi-step tasks. But performance varies across tests: Argon trails several competitors on Terminal Bench 4 and does not top the Intelligence Index.
The advertised price also needs to be read alongside measured usage. Artificial Analysis estimates an Intelligence Index task costs $1.99 at Argon’s promotional rates, compared with $3.26 for GPT-6 Astra. That advantage reflects lower token prices, not lower consumption: Argon used an average of 62,000 output tokens per task in the assessment, against 27,000 for GPT-6 Astra. If the discount ends, the estimated task cost rises to $3.98, based on the report’s figures.
For prospective customers, the comparison is not simply which model has the highest score or lowest per-token rate. Task performance, token consumption, access and final pricing all affect the practical cost and suitability of a model. Argon’s limited rollout also means most potential users cannot yet test those trade-offs directly.
As an affiliate, we earn on qualifying purchases.
A Long Gap Between Frontier Models
Argon is Google’s first frontier model in more than seven months, following Gemini 3.1 Pro. The Decoder reports that Google had previously announced a Gemini 3.5 frontier model but later skipped that release after a difficult, drawn-out development period. Argon’s arrival therefore marks a new competitive entry from Google, though early test results do not establish that it has overtaken its rivals.
The model accepts text, images, video and audio as input and generates text. Google says its input context window is one million tokens and has raised the output limit from 64,000 to one million tokens. A new Gemini API feature, Long Decode Continuation, is intended to pause lengthy responses and resume them through follow-up requests rather than have generation stop at a timeout.
The report also cites a separate result from Vals AI: Argon scored 68.9% on the Vals Index and placed in the top five on 20 of 22 benchmarks. That index showed strengths in finance, law, coding and security, but the source cautions that results should be treated carefully; among other things, the benchmark placed mid-tier Claude Sonnet 5.5 above Anthropic’s top model, Opus 5.5.
““phased approach””
— Google, on its staged release plan, as quoted in The Decoder
As an affiliate, we earn on qualifying purchases.
Public Access and Results Still Pending
Google has not announced when developers, businesses or consumers will be able to use Argon. Its first users are a restricted group of cyber defenders and Google teams, and the scale and duration of that testing have not been stated in the supplied report.
The cited evaluations are early benchmark results, not proof of how Argon will perform across all real-world uses. The source material does not provide enough detail to establish whether results will hold across different prompts, deployments or safety settings. Google’s own benchmark results are described as more favorable than the independent assessment, but the available material does not include a full, independently verified comparison of those claims.
Cost estimates depend on the benchmark’s measured token use and Argon’s introductory rates. Google’s regular listed prices are higher, and the report does not say how long the promotion will last. The results also show a distinction between admitting uncertainty and answering correctly: Argon had a lower hallucination rate on AA-Omniscience, but its accuracy score was 50%, below Gemini 3.1 Pro Preview’s 55% and GPT-6 Astra’s 63% in that test.
As an affiliate, we earn on qualifying purchases.
Testing Before Wider Availability
Google plans to use feedback from Fairwind participants and internal teams to inform Argon’s safety mechanisms. It is also participating in a U.S. government voluntary program that gives agencies access to new models before public release, according to The Decoder.
After this limited stage, Google says it intends to open access first to paying API customers and Google AI Ultra subscribers, followed by broader availability. No firm date has been published. Further independent tests, details about the rollout and confirmation of how long introductory pricing remains in effect will help show whether Argon’s benchmark gains translate into a consistent advantage for users.
As an affiliate, we earn on qualifying purchases.
Key Questions
Does Gemini 4 Argon lead OpenAI and Anthropic’s models overall?
No clear overall lead is established by the results in the report. Argon tied some competitors on Artificial Analysis’s Intelligence Index and led on its AutomationBench-AA test, while Anthropic’s Claude Opus 5.5 and Sonnet 5.5 scored higher on the Intelligence Index.
Who can use Gemini 4 Argon now?
Initial access is limited to selected “trusted cyber defenders” in Google’s Fairwind program and Google’s internal teams. Google has not announced a date for general availability.
How much does Gemini 4 Argon cost?
Google’s introductory API price is $2 per million input tokens and $10 per million output tokens. The report says regular prices are $4 and $20, respectively. Google has not specified how long the introductory rates will last.
Is Argon cheaper to use than competing models?
Its introductory token rates are lower than those cited for several rivals, but that does not always mean a lower cost per task. In Artificial Analysis’s Intelligence Index assessment, Argon used an average of 62,000 output tokens per task, compared with 27,000 for GPT-6 Astra.
What does Google say about Argon’s capacity for long responses?
Google raised Argon’s output limit to one million tokens and introduced Long Decode Continuation in the Gemini API. The feature is designed to pause and resume lengthy responses through follow-up requests.
Source: rss
Columbus Day / Indigenous Peoples' Day Picks
long weekend sales
As an affiliate, we earn on qualifying purchases.
