The gate opened, on schedule nobody published

When Google unveiled Argon on 30 September, the model was not really a product. It was a promise with a guard at the door: approved cybersecurity defenders could use it, everyone else could read about it, and Google would say only that wider access would come as soon as possible. Thirteen days later the guard has stepped aside. Argon is now available to anyone willing to pay, through the API and AI Studio for developers, Gemini Enterprise for companies, and the Gemini app and Search's AI Mode for Pro and Ultra subscribers.

That makes 13 October the first day Argon is broadly accessible to paying developers and eligible subscribers, allowing substantially wider independent testing. Selected external cybersecurity defenders and evaluation organisations had access before the public rollout. Until now every number attached to this model arrived through a controlled channel. From today, enterprise buyers can run it against Claude and GPT-6 on their own work, which is the only comparison that has ever mattered for a model pitched at coding, finance, legal work and research.

Google is separately expanding Fairwind access to vetted cybersecurity defenders, who can use a version without the public model's cyber guardrails.

Two leaderboards, two answers

Google's launch materials lead with the Vals Index, and on that measure Argon earns the lead: 68.90%, first place, ahead of both current Claude models on professional tasks weighted toward finance, coding, legal and tax work. It also tops the Vals finance-agent benchmark.

The other independent measure published alongside the launch is less convenient. On the Artificial Analysis Intelligence Index, Argon scores 53, level with GPT-6 Astra and Claude Fable 5.1, and below Claude Opus 5.5 at 58, Sonnet 5.5 at 56 and Meta Muse Wonder Full, also at 56. Both results can be true at once, because the indexes weigh different work. What they rule out is the simple story. Argon is not clearly the best model available. It is a model that is strongest on a particular kind of sustained professional task, and ordinary on a broader mix. Buyers who quote only the Vals result in an internal proposal are doing Google's marketing for it.

Google reports more than 100,000 monthly internal Antigravity users and, over a four-week period, a 35% increase in agentic code submissions alongside a 30% decrease in post-submission rollbacks. These are measured internal changes as reported by Google, but the company does not disclose the comparison baseline, sample design or controls needed to attribute the improvements solely to Argon. The concrete examples are better evidence: Google says Argon helped refactor Chrome browser-feature code from a roughly six-million-line legacy folder within Chromium's 40-million-line codebase, completing changes across hundreds of files in weeks rather than the months expected; and an overnight Agent Development Kit optimisation cut in-memory session preparation time by 40% and sped cold imports by 21%. Those are specific enough to be wrong, which is what makes them worth more than the percentages.

Price the December model, not the October one

The clearest new fact in this launch is a date. Argon's $2/$10 introductory input/output rates match the published base token rates of Claude Sonnet 5.5 and GPT-6.1 Sol, but cache pricing, context-dependent tiers and other terms differ. From December 1, Google's scheduled $4/$20 rates match the base input/output prices of Claude Opus 5.5 and Meta Muse Wonder. These are list-price comparisons, not equal costs for the same workload. Google did not hide this; it simply announced both numbers at different volumes.

So the honest way to evaluate Argon is at its December price, against Opus-class rivals, with the seven weeks until then treated as a subsidised trial. A team that finishes its testing by late November gets genuinely useful information cheaply. A team that starts in December, or that builds a cost model on the introductory rate and forgets to change it, will pay double for a decision it thought it had already made. Token prices also say nothing about how many reasoning tokens a task actually consumes, which is where comparisons between models are usually won and lost.

One more number deserves a careful read. Google's September announcement described an expanded one-million-token output limit, while its October release and model card describe up to one million input and output tokens combined, enabled by a continuation protocol. Google has not clearly reconciled the two descriptions in its launch materials. Vals' model listing records a 262,144-token maximum output setting for its evaluation, which should not be mistaken for a universal API limit. For the specialised scientific work that might need that kind of budget, the difference matters. For everyone else, it is a reminder to read the current documentation rather than the launch coverage, including ours from 1 October.

Dated note (13 October 2026): access and prices above are as announced on 13 October 2026. The introductory API price of $2 per million input tokens and $10 per million output tokens expires on 30 November 2026; the standard price from 1 December 2026 is $4 and $20. If Google extends the introductory period or changes either price, add a dated update here rather than editing the text above.

Sources

  1. [1] Google, "Gemini 4 Argon is now available for everyone" — October 13 general-availability announcement (access channels, introductory pricing, December 1 standard rates)Read source
  2. [2] Google — September 30 restricted unveiling (Fairwind Program access, voluntary US government pre-release review, original one-million-token output-limit wording)
  3. [3] Google Antigravity — October 13 engineering report (100,000+ monthly internal users, four-week submission and rollback metrics, Chrome refactor and Agent Development Kit examples)
  4. [4] Google DeepMind — Gemini 4 Argon model card (combined input-plus-output token limit via continuation protocol)
  5. [5] Vals — Finance Agent v2 leaderboard and Vals Index (Argon first at 68.90%; Finance Agent v2 at 65.40%, as of 12 October 2026)
  6. [6] VentureBeat, "Google makes Gemini 4 Argon broadly available with limited-time low API pricing of $2/$10 per 1M token", 13 October 2026Read source
  7. [7] Gadgets 360, "Google Gemini 4 Argon Launched Globally for Developers and Enterprises", 13 October 2026Read source
  8. [8] AI Stock Wire, "Google (GOOGL) opens Gemini 4 Argon to developers and paid subscribers", 13 October 2026Read source
  9. [9] AI Affairs, "Google releases Gemini 4 Argon to paying subscribers with lower introductory prices", 13 October 2026Read source