Google's Gemini 4 Argon is GA; $2/$10 price rises to $4/$20 on 1 Dec

Google made Gemini 4 Argon generally available on 13 October 2026, two weeks after it announced the model on 30 September to a limited group of cyber defenders. The Gemini API changelog lists the model ID gemini-4-argon as GA on 13 October. Google's developer blog says developers can use it on the paid tier of the Gemini API in Google AI Studio and Android Studio, and that Google AI Ultra subscribers can use it in AI Studio and in Google Antigravity, its agent development platform.

A separate Google post says AI Pro and Ultra subscribers can use it in the Gemini app. Google's pricing page says it is not available on the API Free tier or the AI Pro plan.

The price changes on 1 December

Google's launch post gave an introductory price of $2 per million input tokens and $10 per million output, and did not say what follows. Its API pricing page does. Standard paid-tier requests cost $2.00 input and $10.00 output through 30 November 2026, then $4.00 and $20.00 from 1 December 2026. Cache reads go from $0.10 to $0.20.

• Batch and Flex: $1 input and $5 output through 30 November, then $2 and $10.
• Priority: $3.60 and $18 through 30 November, then $7.20 and $36.
• Output prices include thinking tokens, and Google says Argon's output limit is up to 1 million tokens, up from 64,000.

The Decoder's price table lists Claude Opus 5.5 at $4 input and $20 output per million tokens. If that is right, Argon's regular price from December equals Opus 5.5's list price, and only the promotional months are cheaper. We did not check Anthropic's price page for this article.

What Google's benchmarks show

Google's model page compares Argon with GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. These are Google-reported results, and the picture is mixed.

• Argon leads on knowledge-work tests: Vals Index 68.9% against 67.0% for Opus 5.5, Vals Finance Agent v2 65.4% against 58.6%, AutomationBench 51.3% against 42.5%, and Harvey's Legal Agent Benchmark 19.6% against 3.8%.
• It leads on DeepSWE v1.1 for coding (77.9% against 74.2% for Opus 5.5), but trails on FrontierSWE v2 (55.0% against 62.3% for Opus 5.5 and 65.5% for GPT-6 Astra) and on Terminal-bench 4.0 (57.4% against 66.4% for Opus 5.5).

Google's own examples are internal: it says Argon agents made a Rust video decoder (libgav1) 2.7 times faster with identical output, and that agents freed more than 300 TiB of memory across its data centres once rolled out.

Safeguards and the Fairwind programme

Google's safety post says Argon solved all 12 simulated end-to-end cyberattack scenarios in its tests, up from five for the previous generation, so it applied its frontier safeguards in cyber and biology: input screening, probes on internal activations, and monitoring that can cut an agent off mid-stream. It is also widening the Fairwind Program for vetted cyber defenders.

What the Hacker News thread said

The Hacker News thread ran past 400 comments. One commenter's objection, with clear basis, was that the launch post gave only the introductory price. The pricing page bears this out: the standard rate doubles on 1 December. Another commenter's image-to-HTML test put Argon below Opus 5.5; that is one individual test, not an evaluation.

What to do

Price any Argon pilot at the December rates, and consider Batch or Flex ($2 and $10 per million tokens from December) for work that can wait. For terminal-style agentic coding, test your own tasks against Opus 5.5 first, since Google's table has Argon behind on two of three coding benchmarks. For legal, finance and long-document work, Google's numbers favour Argon, but they are vendor-run.