Google DeepMind Launches Gemini 4 Argon and Cuts API Prices in Half

Google DeepMind has opened global access to Gemini 4 Argon, the first frontier model in the Gemini 4 family. The API launch comes with a 50% discount for a limited time. The Chinese outlet Jiqizhixin describes the release as long overdue: the lineup took longer to arrive than promised. According to Google, Argon is built for complex real-world tasks, primarily in software development and enterprise knowledge work.

What we know about the model

Google presents Argon as the first model of a new generation. The word "first" matters here: it hints that the family will expand, as earlier Gemini versions did, when lighter variants appeared alongside the flagship. The original announcement does not name the lineup's makeup, so this is only an expectation.

The focus on software engineering and enterprise knowledge is no accident. Agentic coding and work with companies' internal documents now bring in most of the revenue in the API market. Google competes with Anthropic and OpenAI over exactly these use cases, and they are where developers notice the differences between models fastest.

A discount as an entry tool

Half-price API access while the promotion lasts is meant to get teams to test the new model before they settle into their competitors' solutions. It is a standard move, but it works. However, the available material includes neither a base price per million tokens nor an end date for the promotion. There is nothing yet to compare Argon's price against competitors, and budgeting around a promotional rate is risky.

The same goes for quality. Google states its goal, but at the time of publication the source has no figures: no coding benchmark results, no context window size, no response speed.

What can be confirmed right now:

• Model: Gemini 4 Argon, the first frontier model in the Gemini 4 family.
• Developer: Google DeepMind.
• Availability: global launch.
• API price: 50% off for a limited time.
• Stated use cases: software development, enterprise knowledge work.
• Not disclosed in the first announcement: benchmarks, context window, base pricing, promotion duration.

How to approach testing

Until independent measurements appear, your own tasks are the best guide. Teams that already build coding agents or search over a corporate knowledge base should run their real test suites on Argon and compare the results with their current model. The metric here is not price per token but cost per solved task: a more expensive model can end up cheaper if it needs fewer reruns.

While the discount lasts, it is a good time for a cheap experiment. But choose a production model based on the base rate, not the promotional price. If you work with sensitive data, check the processing terms and regional restrictions in Google's enterprise plans in advance. Detailed technical specifications will apparently appear in the coming days, and the picture should then become fuller.