Google is selling its newest flagship model at half its eventual price, and the discount has an expiry date. Gemini 4 Argon, which Google announced on September 30, 2026 and opened more broadly on October 13, costs $2 per million input tokens and $10 per million output tokens through November 30. From December 1, according to Google’s launch post and secondary coverage, the list price doubles to $4 and $20. Any team that wires Argon into a product this month is quietly signing up for a bill that changes on a known date.
A phased release, by design
Argon did not arrive all at once. Google’s announcement said the model was rolling out first to trusted cyber defenders through a program it calls Fairwind, with access widening in phases to paid API customers and Google AI Ultra subscribers after early-tester feedback. Koray Kavukcuoglu, a senior vice president at Google DeepMind, put the logic plainly: “Safely releasing frontier capabilities at this level requires a phased approach.”
The reason is the model’s security skills. Google says Argon can autonomously find, validate and patch serious software vulnerabilities, and that defenders in the Fairwind program get a version without the cyber guardrails the public model carries. Google also says it is taking part in the US government’s voluntary pre-release model access process, and describes safeguards that include misuse refusals, activation monitoring, adversarial training against prompt injection, and monitoring of the model’s reasoning that can halt execution. Those are Google’s own descriptions; none have been independently audited in the material reviewed for this article.
By October 13, coverage described Argon as broadly available: paid Gemini API access for developers, the Gemini app and AI Mode for AI Pro and Ultra subscribers, and Google Antigravity for Ultra. One aggregator reported availability on Vertex AI as well, but other coverage mentioned only the Gemini API and AI Studio, and no Google statement confirming Vertex AI was found. Teams that need Vertex for compliance or billing reasons should check Google Cloud’s own documentation rather than rely on third-party pricing tables.
What the numbers actually say
Google’s headline figures are all self-reported. It cites 77.9% on DeepSWE v1.1, a long-horizon software engineering benchmark; 51.3% on Zapier’s AutomationBench, which tests business workflow tasks; 91.7% on LVBench for long-video understanding; and a tie for first at 68% on CWE-bench v1, a vulnerability remediation test. It also reports that the maximum output length rose to 1 million tokens from 64,000.
That last claim needs a careful reading. According to one developer-focused write-up, Argon works from a single shared budget of 1,048,576 tokens that covers input, reasoning and output together. The default synchronous output is capped at 65,536 tokens, and longer responses go through a “long decode” path with extra requirements, such as background execution on the Interactions API. A long answer is therefore not free: reasoning tokens and repeated input across continuation steps both count, which affects cost as well as latency.
Caching is the other lever. Google says cached input is discounted 95% from the input price, and the same developer write-up puts cached input at $0.10 per million tokens during the introductory period. For applications that resend the same long document or codebase with each request, that discount can matter more than the headline rate.
Budget for December, not October
The practical lesson is to model costs at $4 and $20 from the start. A product that looks comfortable at $2 and $10 may not survive the doubling, and customers rarely accept a surprise price change from a vendor who built on a promotional rate. If the economics only work at the discounted price, that is a signal to test a cheaper model before launch rather than after.
It is also worth separating the benchmark story from the buying decision. Vendor-reported scores on coding, workflow and video tests show where a model is strong, but they say little about how it behaves on a specific product’s data. The reasonable approach is to run a small evaluation on real tasks, measure cost per completed job including retries, and compare against at least one cheaper alternative. The AiCybr write-up that summarized Google’s numbers reported no independent benchmark results for Argon, so the figures above should be treated as claims until third-party testing catches up.
There is also a competitive angle. Frontier labs are now releasing flagship models in tiers, with security-sensitive capabilities gated for vetted users first. That pattern is likely to repeat, which means the version of a model a startup can call in its first week may differ from what trusted partners get. Product plans should not depend on the most powerful variant being publicly available.
What This Means for Philippine Founders
For Filipino teams building on the Gemini API, the first step is to put the December 1 repricing into the financial model now. Most Philippine startups bill customers in pesos while paying model vendors in dollars, so a doubling of token prices combines with exchange-rate movement. Pricing a customer-facing feature at the promotional rate, then absorbing a 2x cost increase, is a risk a small team can avoid with a spreadsheet and an afternoon of testing. Where possible, use the caching discount and keep prompts lean, since reasoning and repeated input drive spend.
It is also sensible to avoid building a business around Argon’s cyber capabilities until access rules are clearer. Google has said broader availability is phased and that the unguarded variant is for vetted defenders; whether and how Philippine security firms or developers could qualify for Fairwind has not been confirmed in the sources reviewed here. Founders who want to work in security tooling should treat that as an open question to raise directly with Google’s sales or partner teams, and in the meantime design products that work with the publicly available model and swap in a cheaper one if the numbers demand it.
Share this article