On September 30, 2026, Google released Gemini 4 Argon, positioned as a frontier model for long-horizon enterprise workflows. It launched first to vetted cybersecurity defenders through the Fairwind program, then expanded to paid API users and Google AI Ultra subscribers.
The pricing is the most aggressive part of the release. Introductory rates: $2 per million input tokens, $10 per million output, with a 95% cache discount. Standard pricing takes effect January 1, 2027: $4 input, $20 output. The single-response output ceiling rises from 64K to 1 million tokens.
Google's published benchmarks place Argon at 77.9% on DeepSWE v1.1, first on the Vals enterprise knowledge work index at 68.9%, and first on AutomationBench at 51.3%.
The Public Model Is a Rebadged Internal Version
Business Insider reported that Google internally tested at least three Gemini 4 derivatives: Argon, Barium, and Carbon.
Internal documents show that the model released publicly as “Argon” is internally designated “Barium-B.” The version customers received was one candidate selected from several.
Carbon is newer. It recently entered Jetski, Google's internal coding platform, for employee testing. One employee described the coding experience as feeling “like Opus 5.5.” Another wrote in an internal channel that “Carbon is really good.”
Whether Carbon ships as an update to Argon or as a separate Gemini 4 model is undetermined. Google declined to comment.
Independent Testing Shows a Different Picture
Artificial Analysis's independent evaluation diverges from Google's launch materials.
On the AA Intelligence Index, Argon scores 53, level with GPT-6 Astra and one point ahead of GPT-6.1 Sol. Its advantage comes from a lower hallucination rate: Argon acknowledges it doesn't know the answer 42% of the time, versus 19% for Astra. But Argon's correct answer rate is also lower: 50% versus 63%.
On Terminal Bench 4, Argon scores 57%, below Claude Sonnet 5.5 at 64%, Claude Opus 5.5 at 60%, and GPT-6 Astra at 59%.
On cost, introductory pricing brings Argon's per-task cost to $1.99, about 60% of Astra's. But the advantage comes from the price, not fewer tokens: Argon generates roughly 62,000 output tokens per task, versus 27,000 for Astra. When standard pricing takes effect, per-task cost rises to $3.98 — about 1.2 times Astra's.
A Million Tokens of Output Cuts Both Ways
Argon's 1 million token output ceiling means a model can reason through a single task trajectory and produce hundreds of thousands of tokens of output. That enables large-scale code migration, deep research, and multi-step enterprise workflows.
It also amplifies the cost of error. An early misjudgment that goes uncorrected can have subsequent edits, tests, and explanations built on a faulty premise. Evaluating next-generation coding models requires more than measuring how long they can work. It requires measuring when they recognize a wrong direction, stop unproductive attempts, and hand uncertainty back to the user.
P.S. Google has not said whether Carbon will be released publicly or explained the differences between the public Argon and internal versions. Fairwind partners are using Argon. Google employees are using Carbon on Jetski. The gap between them can currently only be inferred from internal employee descriptions.
Frequently Asked Questions
Q: What did Google release?
A: Gemini 4 Argon on September 30, 2026. It launched to vetted cybersecurity defenders through Fairwind, then to paid API users and Google AI Ultra subscribers.
Q: What is the pricing?
A: Introductory: $2 per million input tokens, $10 per million output, 95% cache discount. Standard pricing from January 1, 2027: $4 input, $20 output.
Q: What is the Barium-B finding?
A: Business Insider reported that internal documents designate the publicly released Argon as "Barium-B." Carbon is a newer internal checkpoint being tested by employees on Jetski.
Q: How does Argon perform on independent benchmarks?
A: It scores 53 on the AA Intelligence Index, 57% on Terminal Bench 4, and $1.99 per task at introductory pricing. It generates roughly 62,000 output tokens per task versus 27,000 for GPT-6 Astra.
Q: What has Google not said?
A: Whether Carbon will be released, and what differs between the public Argon and the internal versions.
