Google's Gemini 4 Argon Ties Astra. The Discount Has a Deadline.
Gemini 4 Argon matches Astra on an independent index at $2/$10 per million tokens. The price doubles in December, and tokens are only half the story.
Google has an AI model matching GPT-6 Astra on a major independent benchmark, and the launch price is $2 per million input tokens and $10 per million output tokens. Your API bill would like a word. So would the expiry date.
On October 13, Google DeepMind opened Gemini 4 Argon to developers and enterprises through AI Studio and Google's cloud platform. Google AI Pro and Ultra subscribers also got access in the Gemini app and AI Mode in Search. The model that had been behind a restricted rollout finally had a much bigger front door.
The Astra comparison comes from Artificial Analysis. Argon at its high reasoning setting scores 53 on the Intelligence Index, matching GPT-6 Astra at max. That's a tie on that index, with those settings. Your enormous, slightly haunted codebase has not signed off on the result.
Google is pitching Argon for work that takes more than one clever answer: substantial code changes, financial analysis, legal work and scientific research. Its launch demos include building 3D scenes and games through repeated rounds of generating, inspecting and revising. Less autocomplete, more intern who has discovered Blender and refuses to go home.
Then comes the calendar. The $2 input and $10 output prices last through November 30, 2026. On December 1, they become $4 and $20. Cached input is cheaper too: 10 cents per million during the offer, then 20 cents. Google has put the introductory price in very large type. Your spreadsheet should give the end date equal billing.
Cheap tokens also have a habit of inviting more tokens. In Artificial Analysis's evaluation, Argon averaged roughly 62,000 output tokens per task, compared with Astra's 27,000. At the discounted rates, an Intelligence Index task cost $1.99 on Argon versus $3.26 on Astra. A useful saving. A much chattier way to get there.
At standard pricing, Artificial Analysis projects Argon's cost for that same evaluation at $3.98 per task. The bargain changes sides. Those are benchmark costs, not a quote for your application, but they make the point rather efficiently for two models that enjoy thinking at length.
Developers get a unified window of about one million tokens, shared by input, thinking and output, with controls for how deeply the model reasons. There is no free API tier for Argon. Bring a billing account, and remember that thinking tokens count toward output charges. The model's internal monologue also has a meter.
The useful move is to give it work you can actually judge: a familiar bug, a document you know well, a repetitive task whose finished result matters more than the demo. Count the retries, check the answer, and price the run twice. Once for the introductory offer. Once for December.
Google has made a credible case for a place in your model picker. Whether it earns a permanent spot depends on what comes back and what it costs to get there.
Would Argon make it into your daily workflow, or is this a trial relationship that ends when the discount does?