Gemini 4 Argon goes GA with 50%-off pricing

Google made Gemini 4 Argon generally available on October 13, ending the two-week invite-only phase for the flagship model it unveiled on September 30. The paid API is now open to everyone through AI Studio and Android Studio, and the Gemini app serves AI Pro ($20/month) and AI Ultra (from $99/month) subscribers. The number that matters most in the announcement is a date: intro pricing of $2 per million input tokens and $10 per million output tokens holds until November 30, when rates double to $4/$20.

Key points

Argon is out of its invite-only phase: paid API (via AI Studio and Android Studio), plus the Gemini app and AI Mode in Search for AI Pro and AI Ultra subscribers, per Google’s October 13 post.
Intro pricing of $2/$10 per million tokens, with cached input at 95% off the input rate, expires November 30; regular rates are $4/$20.
Inside Antigravity, only AI Ultra subscribers get Argon, and Ars Technica reports Google declined to say when — or whether — other tiers will get it.

Background

At launch on September 30, Argon was a benchmark sheet more than a product: the benchmarks and the intro pricing ($2/$10 per million tokens) were public from day one, but there was no general-availability date and no end date for the promo — our entry that day led with exactly that gap. The New Stack’s headline caught the mood: “Gemini 4 Argon is here: It’s great, and you can’t have it yet.” During the invite phase, access went to cyber defenders through Google’s Fairwind program (Wiz among the first); the GA post adds “hundreds of additional” Fairwind partners. Google is also not alone in splitting availability from announcement — OpenAI halted the already-announced GPT-6.1 last month — so flagship labs are now treating “you can use it” as a separately scheduled event.

The facts

Every number below traces to Google’s GA post or Ars Technica’s report; the pricing structure is the part budget owners should copy into their spreadsheets.

Tier	Intro (through Nov 30)	Regular (after Nov 30)
Input	$2 / M tokens	$4 / M tokens
Output	$10 / M tokens	$20 / M tokens
Cached input	95% off the input rate	not announced

Context: a 1M unified input-output token limit via the Gemini API — the “1M output” from launch day is now one shared budget across both directions.
Availability: developers build on the paid API in AI Studio and Android Studio; consumers get the Gemini app and AI Mode in Search on AI Pro and AI Ultra; Antigravity, and in-app Argon in AI Studio, are Ultra-only; enterprises get “Gemini for business”, the API and Antigravity. Vertex AI is never mentioned in the post.
Benchmarks (official claims): first on the Vals Index, which covers finance, coding, legal and tax; 87% on the embodied-reasoning benchmark ERQA, which Google says beats all other frontier models; a 0.7% attack success rate on Gray Swan’s indirect prompt-injection benchmark.
Scientist program: 10,000 scientists can apply for six months of the highest Antigravity usage tier free of charge; onboarding is rolling and applications are open.
Internal use (self-reported): agentic code submissions in Antigravity up more than 35%, rollbacks down by nearly a third.

The price-performance math

Artificial Analysis’ independent numbers give the promo price a reference point. Argon scores 53 on its Intelligence Index — rank 8 of 216, against a peer-category median of 27 — exactly tied with GPT-6 Astra. Cost per task is $1.99 for Argon versus $3.26 for Astra, so equal intelligence at roughly 60% of the cost. On a blended cache/input/output price (weighted 7:2:1), Argon runs $1.47 per million tokens against Astra’s $7.70. The benchmarks carry one known weak spot: the first external evaluations, which our September 30 entry cited, put Argon’s AA-Omniscience hallucination rate at 15% versus Opus 5.5’s 59%. Low for that batch — but those runs happened during the invite phase, no third party has re-tested the GA build yet, and your own eval set is still the one that counts.

What others say

Ars Technica framed its report around the tier wall. AI Pro subscribers ($20/month) can write code with Argon in the Gemini app’s Canvas mode but cannot enter Antigravity; Google’s explanation is that Antigravity’s multi-threaded agentic workflows are “probably too expensive”, and on when other tiers get access, the company declined to comment. Free users sit further out still: Ars notes Google recently cut the free tier’s Flash models back to Flash-Lite. The caution is Google’s own framing — the model is trained to refuse malicious instructions, carries an automated misalignment monitor, and posts that 0.7% prompt-injection figure. The New Stack’s September 30 headline stood for how everyone felt two weeks ago: great model, you can’t have it. As of October 13, the second half of that sentence has expired for paying customers.

Our take

The line to copy into your calendar is the deadline. Six weeks of $2/$10, then $4/$20, amounts to an official migration window: run your proofs of concept inside it, negotiate contracts against the post-November-30 rate, and keep the two ledgers separate. The math justifies shortlisting Argon: equal intelligence to GPT-6 Astra at about 60% of the per-task cost undercuts even GPT-6.1 Sol’s budget positioning right at the flagship layer. Two reservations. First, Antigravity is Ultra-only, and Google has not committed to opening it to Pro — agentic-heavy teams should decide whether the $99/month cover charge is worth it before building on it. Second, Vertex AI never appears in the post, so procurement paths that run through GCP’s enterprise pipeline are still waiting on a word. And the 15% hallucination rate has not been overturned by newer data; test long-document retrieval workloads before you commit.