Gemini 4 Argon Goes Public at $2/$10, Antigravity Locked to Ultra

Argon is out, with a price that expires on November 30

Google has opened its first Gemini 4 model to the public, but only to customers who pay, and its launch price will not last. On October 13, 2026, Gemini 4 Argon started rolling out globally to Google AI Pro and Ultra subscribers in the Gemini app.14 Google also made it available in AI Mode in Search, in the Antigravity coding environment for Ultra subscribers, and to developers and enterprises through the Gemini API in Google AI Studio and the Gemini platform.4 The public launch came 13 days after Google announced the model and limited it to a small group of vetted cybersecurity partners.9

The API starts at $2 per million input tokens and $10 per million output tokens, and cached input is discounted 95%.4 Those rates end on November 30, 2026. From December 1, the price doubles to $4 per million input tokens and $20 per million output tokens.2 VentureBeat notes that Google's September 30 announcement gave no end date for the discount. The October 13 launch is when the cutoff became official.2

The story has three parts: a staged release that runs through cyber defenders first, a subscription split between chat and coding, and pricing that looks very different depending on how cost is measured. Taken together, they show Google is trying to win back attention after a difficult year without giving away compute on the products that cost the most to run.

How Google released it

Argon did not go straight to the public. When Google announced it on September 30, access was limited to cyber partners in its Fairwind Program. Google said the model was trained for defensive security work and could find, validate and patch software vulnerabilities on its own.13 The New York Times connected that caution to a wider industry debate about AI safety. It reported that Argon would reach more users only after defenders had time to fix the bugs it found, and that the model shipped with monitoring that can catch and stop it if it goes beyond what a user intended.19 CNET described this as the new standard order for frontier releases: cybersecurity experts first, then paying subscribers, then free users.17

The two-week gap was short, and Google's stated reasons varied by outlet. Google Cloud says Argon blocks offensive tasks such as pentesting and exploitation by default. A separate version without those guardrails is available to Fairwind members, including government, critical infrastructure and security partners.18 Google also calls Argon its most resistant model yet to indirect prompt injection. It reports that Argon cut attack success to 0.7% across 15 adaptive attempts on the Gray Swan benchmark.18 These are Google's own numbers, and at least one outlet stresses that the published results have not been independently measured.10

The real story is who can use it where

Nearly every outlet agrees that Argon is paid-only and that free accounts are excluded.1015 The more important detail is which paid plan unlocks which product. Pro subscribers, who pay about $20 a month, get Argon in the Gemini app and AI Mode, and can generate code in Canvas. Using Argon in Antigravity or AI Studio requires Ultra, which starts around $99 a month.510 Ars Technica and others report that Google has not said whether Argon will reach other Antigravity tiers, or when.10 BigGo reports that Google sees Antigravity's multi-threaded agent workflows as too compute-heavy for lower tiers.5

Users have noticed. One reply to Google's launch post complained that Pro users had no Argon in Antigravity and said the writer would have canceled if not locked into a yearly plan.4 The split matters because coding is the area Google most needs to win. Sundar Pichai said on the July 2026 earnings call that coding and agentic coding were areas where Google needed to improve.11 Google says Argon already handles much of its internal engineering work, and the update also puts Antigravity's agent coding interface inside the Gemini app on macOS.3

In our reading, putting Antigravity behind Ultra is a capacity decision presented as a perk. Google is willing to show Argon to Pro users in chat but not pay for long agent runs at the $20 tier. Teams that want Argon for coding without an Ultra subscription have other options. Lindy reports it is available in four GitHub Copilot plans (Pro+, Max, Business and Enterprise) at Google's list price, and Google Cloud lists it in Gemini for business.618

The exact Ultra price varies by outlet. Figures include $99, $99.99 and $100 for the entry tier. Mashable notes that Google added a $100-a-month Ultra plan at I/O 2026 and cut the top tier from $250 to $200.367 The figures differ by rounding, not substance.

Cheap per token, less cheap per task

At introductory rates, Argon costs half as much per token as Claude Opus 5.5 and one-fifth as much as OpenAI's GPT-6 Astra, according to VentureBeat.2 Lindy's cost comparison puts Argon at the same $2/$10 token price as GPT-6.1 Sol. After December 1, Argon's price matches Claude Opus 5.5 at $4/$20, while GPT-6 Astra stays far higher at $10/$50.6

Per-token price is only part of the cost. Benchmarking firm Artificial Analysis gave Argon a 53 on its Intelligence Index, the same as GPT-6 Astra. Argon's cost per task at introductory pricing is $1.99, about 60% of Astra's cost.5 Mashable reports that once the discount ends, Argon's cost per task rises to $3.98, about 1.2 times Astra's.3 Forbes cites an eesel analysis showing GPT-6.1 Sol completing the same work for $0.72 per task at the same token price, because Argon produces about 62,000 output tokens per task.9

This is where the coverage disagrees most, and we think Forbes's framing is the right one. Argon's price advantage is real but temporary, and it depends on how much text the model generates. Argon is a reasoning-heavy model, and output tokens (including its thinking tokens) are billed at the higher rate.6 Once the discount ends, it is no longer clearly the low-cost option against OpenAI's most expensive model. One point in Argon's favor: BigGo, citing Artificial Analysis, reports a 15% hallucination rate for Argon, compared with 51% for GPT-6 Astra and 54% for GPT-6.1 Sol.5 That gap could matter more than price for regulated legal and finance work.

Context window and output limits

The Argon API supports a context window of about 1 million tokens, but input and output now share that limit.68 Lindy lists default output at 65,536 tokens, with up to about 1 million when run as a background job. It also lists a March 2026 knowledge cutoff and three thinking levels.6 Some coverage describes Argon simply as having a "1-million-token output limit," which leaves out that this applies to long background jobs.17 Fello AI notes that earlier Gemini models were capped at about 64,000 output tokens per response.11

Google had to answer OpenAI and Anthropic

Argon arrives after months in which Google appeared to fall behind. It was Google's first major model since February, and the Times describes a company whose models had lost ground to rivals, especially in coding.19 Google discussed Gemini 3.5 Pro but never released it. DeepMind chief Koray Kavukcuoglu said the company "took a little bit of a step back" to focus on Flash models.12 His comments about Gemini 4 being "as soon as possible" came days after Anthropic and OpenAI shipped major updates to their flagship models.12

Google's launch materials compare Argon with OpenAI's GPT-6 Astra and Anthropic's Fable and Opus models, and cite Vals AI to claim the top spot on its index.1319 Some caution is warranted. Bloomberg reported internal concerns in the final week of September that Gemini 4 was not yet strong enough on tasks such as coding, according to CNET.17 Mashable describes Argon's coding results as mixed, with its clearest leads in knowledge work.3 Business Insider reports that Google tested internal Gemini 4 checkpoints named Argon, Barium and Carbon. One employee said Carbon "feels like Opus 5.5" for coding, but there is no guarantee it will ship.11

The release plan points the same way. Product lead Tulsee Doshi told Business Insider that Argon is the first of a family of new models. She said the public version is the one Google benchmarked in September, even though employees are already testing improved versions.15 Google's own posts say it looks forward to releasing the rest of the Gemini 4 family.411

The bottom line

One point is missing from the coverage: there is no open-weight version of Argon. Google's newest frontier model is available only through paid APIs and subscriptions. Its most expensive coding product is limited to the top tier, and its launch discount ends in about seven weeks.

The outlets broadly agree on the basic facts and disagree mainly on how to read the pricing. Our view is that Argon brings Google back to frontier level on knowledge work and reliability. Whether that holds for coding is still unproven. Developers should test it before November 30, measure their cost per task rather than per token, and plan for the next Gemini 4 checkpoint, which Google itself signals is coming.