On October 13, Google moved Gemini 4 Argon beyond restricted access and into paid developer and consumer products. The wider rollout gives people building with AI another system to consider for work that takes many steps, from reorganizing software to making sense of documents. Google’s announcement makes access the immediate news: Argon was first introduced on September 30.
The important shift is from answering a question to staying with a task. That is the ambition, rather than a guarantee that an agent can finish every assignment unsupervised.
Who can use Gemini 4 Argon?
Developers and enterprises can access the paid API through AI Studio and Google’s Gemini platform. Google AI Pro and Ultra subscribers get Argon in the Gemini app and AI Mode in Search; access through the Antigravity coding environment is for Ultra subscribers. These are different access routes, not a free public release.
The half-price window has an end
The introductory API rates are $2 per million input tokens and $10 per million output tokens, half the stated standard rates. Tokens are the units used to meter model input and output. Google now specifies that the promotion expires November 30, 2026, after which the public-access model costs $4 and $20 respectively.
That discount makes experimentation cheaper. It does not establish which model will deliver a completed job at the lowest cost.
Cheap tokens do not tell the whole story
Artificial Analysis’s September 30 evaluation helps explain why. At high reasoning, Argon scored 53 on its Intelligence Index, matching GPT-6 Astra at maximum reasoning in that comparison. At promotional prices, Argon cost $1.99 per index task versus $3.26 for Astra. But Argon used about 62,000 output tokens per task, compared with 27,000 for Astra. Its lower token prices outweighed heavier consumption.
Those are results from a particular test suite and model settings, not a universal ranking or a quote for running someone’s business. The evaluation predates broad availability; its statement that Argon was restricted describes September 30, not today’s access.
Google’s own model card also shows an uneven frontier. Argon scores above the listed competitors on some knowledge-work tests, but trails GPT-6 Astra and Claude Opus 5.5 on FrontierSWE v2, a software-engineering benchmark. Google acknowledges hallucinations and occasional slowness or timeouts.
For anyone choosing an AI agent, our takeaway is practical: compare the cost and quality of a finished task, including corrections and review. The opening offer changes the price of trying Argon. What matters after that is how much usable work survives the experiment.
