What it is
Gemini 4 Argon is Google DeepMind's flagship reasoning model, announced on 30 September 2026 and first given to vetted cyber defenders in Google's Fairwind programme. Rollout to Google AI Pro and Ultra subscribers began on 13 October, and the model is now listed as available in the Gemini API. Google aims it at software engineering, maths and science, and multi-step agents.
It reads text, images, video, audio and PDF files and writes text. Input and output share a window of 1,048,576 tokens; a single response is capped at 65,536 tokens by default, and a long-decode mode stretches output towards the full million. Thinking has low, medium and high levels.
API pricing is an introductory $2 per million input tokens and $10 per million output tokens (thinking included) until 30 November 2026, then $4 and $20. Batch and Flex run at half price. There is no free tier, and only implicit caching is supported.
Our review
Argon is Google's first flagship since the Gemini 3 series in November 2025, and it puts the company back among the top labs rather than ahead of them. Artificial Analysis scores it 53 on its Intelligence Index at high thinking, level with GPT-6 Astra (max) and Claude Fable 5.1, and five points behind Claude Opus 5.5 at max effort. Google's own launch table is far rosier: per VentureBeat it claims 12 wins out of 18 tests against Astra and Opus 5.5, including 84.2% on the GraphWalks long-context test against Opus's 66.8% and 91.7% on LVBench long video.
The losses in that same table are telling. Opus 5.5 leads on Terminal-Bench 4.0 (66.4% against 57.4%) and Astra on FrontierSWE v2 (65.5% against 55.0%), so terminal-driven coding agents are not Argon's strength. Tech Times adds that every independent score uses high thinking, while the standard model starts on medium. Where it does stand out independently is restraint: a 15% hallucination rate on Artificial Analysis's Omniscience test, the lowest of any model scoring above 45, against 51% for Astra.
It is slow to start. Artificial Analysis measures 45.6 seconds to first token and about 79 tokens a second, and it spent 110 million output tokens on the index, more than most rivals. Because thinking is billed as output, that verbosity turns into money: about $1.99 per index task at the introductory price, roughly double from December.
The price rise is the trade-off buyers need to plan around. At $4/$20 Argon costs the same as Opus 5.5 per token while using more tokens per task, and explicit caching, which makes long shared prompts cheap on Claude, is not offered. What it buys instead is input breadth: an hour of recorded meeting or a stack of scanned PDFs goes in natively, with no transcription step.
Where it earns its place
A media-monitoring team processing several hundred hours of broadcast a week
Pull claims, speakers and timestamps out of long video without a separate transcription pipeline
Send the video file straight to the Gemini API and ask for structured output. Google reports 91.7% on LVBench long video, ahead of Astra and Opus 5.5; running non-urgent jobs through Batch halves the bill.
A legal-tech engineer building review over 500-page contract bundles
Answer questions across a whole data room in one prompt
The 1M-token window holds the bundle, and Google's 84.2% on GraphWalks long context suggests it keeps track across it. At December pricing a 200K-token prompt costs $0.80 before output, and without explicit caching every new question pays most of that again unless implicit caching hits.
A head of engineering at a 200-person company already on Google Cloud billing
Decide whether to standardise on Argon or Claude for internal agents
Argon is cheaper per token until 30 November, then matches Opus 5.5 at $4/$20 while using more output tokens per task. For terminal-heavy coding agents the published Terminal-Bench gap favours Opus; for document and media work Argon's input range is the deciding factor.
Before you commit
- Setup
- Model ID gemini-4-argon in the Gemini API; a Cloud project with billing is required because there is no free tier. Thinking level is set per request (low, medium or high).
- Integrations
- Gemini API with Batch, Flex and Priority inference; Google Search and Maps grounding (5,000 free requests a month shared across Gemini 3 and later, then $14 per 1,000); computer use in preview.
- Data and privacy
- On paid tiers Google says prompts and responses are not used to improve its products, but it logs them for 55 days for abuse monitoring; grounded requests are kept 30 days. The invite-only Contributor tier does use data. No zero-retention option is published on the terms or policy pages.
- Cost at scale
- 10,000 document reviews a month at 200K tokens in and 4K out costs about $4,400 at the introductory price and $8,800 from December; Batch halves either figure. Thinking tokens bill as output, and Artificial Analysis counts about $1.99 per index task now, roughly $3.98 after the rise.
- Limits
- 1,048,576 tokens shared between input, thinking and output; 65,536 output tokens per response by default (32,768 for non-streaming HTTP, per Tech Times). Per-model RPM and TPM are shown only in AI Studio, with spend caps from $250 at Tier 1 to $20,000 at Tier 3.
What's good
- Native video, audio, image and PDF input in one 1M-token window
- Lowest hallucination rate among top-scoring models on Artificial Analysis
- Batch and Flex at half price
- Paid-tier API data is not used for training
What's not
- Price doubles to $4/$20 on 1 December 2026
- 45-second median wait before the first token
- Trails Opus 5.5 on Terminal-Bench 4.0 and Astra on FrontierSWE v2
- No explicit prompt caching; no free tier
Best for
Teams working with video, audio or very large document collections who want a top-tier reasoning model.
Alternatives
Gemini 4 Argon: common questions
- Is Gemini 4 Argon worth it?
- Google is back at the frontier with Argon, and its native video and audio input is something the Claude and GPT flagships still cannot match. We rate it 4.5 out of 5.
- How much does Gemini 4 Argon cost?
- API: $2 input / $10 output per 1M tokens until 30 Nov 2026, then $4 / $20; Batch and Flex half price; cached input $0.10 (then $0.20). No free tier.
- What are the downsides of Gemini 4 Argon?
- Price doubles to $4/$20 on 1 December 2026; 45-second median wait before the first token; trails Opus 5.5 on Terminal-Bench 4.0 and Astra on FrontierSWE v2; no explicit prompt caching; no free tier.
- Who is Gemini 4 Argon best for?
- Teams working with video, audio or very large document collections who want a top-tier reasoning model.
- How much does Gemini 4 Argon cost for a team or in production?
- 10,000 document reviews a month at 200K tokens in and 4K out costs about $4,400 at the introductory price and $8,800 from December; Batch halves either figure. Thinking tokens bill as output, and Artificial Analysis counts about $1.99 per index task now, roughly $3.98 after the rise.
- Is Gemini 4 Argon safe for company data?
- On paid tiers Google says prompts and responses are not used to improve its products, but it logs them for 55 days for abuse monitoring; grounded requests are kept 30 days. The invite-only Contributor tier does use data. No zero-retention option is published on the terms or policy pages.
- What are the best Gemini 4 Argon alternatives?
- The closest alternatives we have reviewed are GPT-6 Astra, Claude Opus 5.5, Claude Fable 5.1 and Gemini 3.8 Flash.
Answers checked Oct 22, 2026.
Make Gemini 4 Argon? Show this rating on your site
