Google's Gemini 4 Argon Is Priced at a Fifth of GPT-6. Almost Nobody Can Use It Yet.

We build agentic products on top of frontier model APIs, so model pricing is a line item in our architecture, not industry gossip. On September 30, Google DeepMind unveiled Gemini 4 Argon, its first flagship update since Gemini 3, a gap of nearly a year, per Axios. The headline is price. The catch is that most of the market can’t buy it yet.

The numbers

Across the 18 disclosed benchmarks outlets have compiled, Argon leads or ties on 13, per coverage from VentureBeat and The Decoder:

Legal analysis (Harvey’s benchmark): a 19.6-point margin over the next-best model. Gaps between frontier models on this test are usually 3–5 points.
Business automation (AutomationBench): 51.3%, leading.
Software engineering (DeepSWE): 77.9%, leading.
Hardest software engineering (FrontierSWE v2): 55%, versus 65.5% for OpenAI’s GPT-6 Astra. Argon trails significantly here.
Output ceiling: 1 million output tokens, up from the previous generation’s 64,000. That matters for long code generations, large document rewrites, and agent steps that need to emit a lot in one pass.

Pricing, per Benzinga:

Introductory: $2 per million input tokens, $10 per million output tokens. Roughly one-fifth of GPT-6 Astra’s launch pricing.
Standard, after the intro period: $4 input, $20 output, matching Anthropic’s Claude Opus 5.5.

Note what that means: the “one-fifth” price is the introductory price. The standard price is level with Opus 5.5, not below it.

Why you can’t just switch

Google chose a staged rollout. Argon went first to a small group of trusted partners through its “Fairwind Program,” aimed at cybersecurity defenders, per SiliconANGLE. It is also going through the U.S. government’s voluntary pre-release review. Availability for paid API customers and Google AI Ultra subscribers is promised “as soon as possible,” with no firm date.

So the benchmark and pricing claims can’t be independently tested by most of the ecosystem. “The race remains unsettled,” as one outlet put it. Every outlet the briefing covers repeats that caveat.

Parity, not dominance

Coverage splits on how much this changes:

VentureBeat and The Decoder say Google is closing the gap and retaking the lead in several categories. The Decoder is explicit that Argon “doesn’t take a clear lead” overall.
CNBC argues that benchmark leadership is necessary but not sufficient. Enterprises buy on reliability, tooling and existing integration, where OpenAI and Anthropic have more real-world mileage.

Timing is part of the story. Per Axios, OpenAI reportedly scrapped a version of GPT-6.1 Astra over security issues before DevDay, and both OpenAI and Anthropic are reportedly investigating tens of thousands of agent-misbehavior incidents. OpenAI’s answer on price came the same week: GPT-6.1 Sol, pitched as near-Astra intelligence for a fifth of the price, per CNBC’s DevDay recap. Two labs are now claiming roughly the same price point, and neither claim is something most buyers can verify yet.

What it means for anyone building on these APIs

We’re not going to call a winner. What the reporting supports:

The “two-horse race” framing is out of date. A benchmark-competitive model at a fraction of the price gives enterprise buyers negotiating leverage they haven’t had in a while.
Benchmarks and real workloads diverge, and Argon’s one significant deficit is on the hardest software-engineering benchmark in the set (FrontierSWE v2, by 10.5 points).
Pricing built into your unit economics today should assume the intro rate is temporary.

Open questions we’re watching: when Argon reaches general API availability, and whether performance holds outside curated benchmarks; whether the government pre-release review becomes a template other labs adopt or a political flashpoint; and how OpenAI and Anthropic respond on price in the next few weeks.