Google opens Gemini 4 Argon more widely. The discount has a deadline.

The same model, different doors

Google’s September introduction began with trusted cyber defenders in its Fairwind programme. The wider rollout now reaches the paid API, consumer apps and Antigravity, Google’s agent-based development environment. DeepMind’s original announcement marks that change.

A Pro subscription gets Argon in the Gemini app, but Google names Ultra for Antigravity. Its personal-account help page also lists Argon for Pro and Ultra tasks; work and school accounts depend on their Workspace licence. Access follows the product, not simply the model name.

The price today is not the price in December

The API pricing page makes the promotion’s end explicit: standard input and output rates double on 1 December. Thinking tokens—the model’s reasoning before its final answer—are billed at the output rate. A short final answer can therefore conceal a much longer paid computation.

Cached input is cheaper when the service can reuse earlier input, but Argon supports implicit caching only: developers cannot create an explicit reusable cache for it. Search grounding has separate charges beyond the stated allowance. The table isolates token rates so those extras do not disappear into a headline price.

A leading score, with a very specific meaning

Vals AI independently places Argon first on its Vals Index. That combines eight evaluations across finance, coding, legal and tax work, weighted by those sectors’ shares of US GDP. It measures benchmark performance, not observed economic gains—and its US weighting is no direct measure of usefulness to a European business.

Vals’ Argon configuration uses high reasoning. Some Claude comparison scores include fallback models after provider refusals; counting those tasks as failures lowers Sonnet and Opus to 65.85% and 65.05%, respectively, without changing their order. A combined lead also does not mean winning every component.

Less guessing does not mean knowing every answer

Artificial Analysis’ September evaluation gives Argon, at high reasoning with Long Decode Continuation, an Intelligence Index score of 53, matching GPT-6 Astra at maximum effort. Its AA-Omniscience test finds a 15% hallucination rate for Argon, but 50% accuracy, against Astra’s 63%. Argon’s advantage there is greater willingness to acknowledge uncertainty, rather than getting more questions right.

The same evaluation puts Argon’s discounted cost at $1.99 per index task, rising to $3.98 at standard prices. It reports about 62,000 output tokens per task against Astra’s 27,000: cheaper tokens do not automatically mean less work. These are that evaluation’s September conditions, not a fresh head-to-head rerun of the October rollout.

A useful production example—and whose example it is

George Kenwright, a creative producer at Google DeepMind, supplies a more tangible example than a leaderboard. He says he gave Argon two green-screen stills, asked it to build a room in Blender, match both camera angles, use Nano Banana to texture the background, and composite it into the original video.

His post shows the stages and the two resulting views. It is a Google-affiliated demonstration involving several tools; it does not establish a typical completion time, price or success rate. The interesting production idea is maintaining one scene across cameras, rather than generating unrelated attractive backgrounds.

The million-token window has one shared budget

The API specifications accept text, images, audio, video and PDFs, but return text; Argon itself does not generate images. Its 1,048,576-token window is shared by input, thinking and output. It is not a million tokens of documents plus another million of reasoning.

Google’s long-decode guidance explains how responses can exceed the default 65,536-token limit: background execution runs continuation steps, with an explicit larger output budget. The output cap includes thinking. Medium effort is the default; high uses more tokens and takes longer to begin answering.

For anyone evaluating Argon now, that makes two comparisons worth keeping together: the quality of the finished work and the bill after the promotion. Run the same task at different thinking levels, inspect the output, and budget against the December rate before deciding what belongs in regular use.