Available now
Trusted cyber defenders
Selected partners in Google's Fairwind Program receive Argon first, and receive it without cyber guardrails.
Announced September 30, 2026 · Rolling out in stages
Gemini 4 Argonlong research, written out in full
Google's first frontier model since Gemini 3, built for deep reasoning across long, multi-step work. It lifts the output ceiling from 64K tokens to 1M, and it leads Google's published numbers on knowledge work rather than on coding.
The Felo API is in closed beta — Gemini 4 Argon access is coming soon.
- 1M
- Max output tokens
- Up from 64K — a full report generated in a single trajectory.
- $2 / $10
- Introductory API price
- Per million input and output tokens. Cached input is 95% off.
- 77.9%
- DeepSWE v1.1
- Google's reported score, ahead of Claude Opus 5.5 and GPT-6 Astra.
- 15%
- Hallucination rate
- The lowest measured among frontier models by Artificial Analysis.
Figures and charts on this page come from the organisations that published them — Google, Arena.ai, and Artificial Analysis — and are reproduced unmodified.
Rollout status
Argon is not generally available yet. Here is the order it arrives in.
Google is releasing Argon in stages, and the first stage is not the public one. Access inside Felo is being set up alongside this rollout.
Announced next
Paid API customers and Google AI Ultra
Google says wider access begins here. No date has been given, and a subscription alone does not grant it today.
Not yet scheduled
Developers, enterprises, and consumers
Follows the paid tiers. Google has not committed to a timeline for general availability.
Argon access inside Felo Search is in progress. This page will be updated when it opens.
What it is good at
A model built to finish long work, not to answer fast
Argon's published strengths cluster around reading, reasoning, and writing at length. That is a different shape from the coding-first frontier models it is being compared to.
Reasoning that survives a long task
Google positions Argon for multi-step work that runs for hours, where earlier models drift or stop early.
1M tokens of output in one pass
Google's stated reason for the increase: a model with room to think at length can solve a hard problem in one trajectory instead of many round trips.
Knowledge work leads, not coding
The widest reported gaps are on Harvey's Legal Agent Benchmark and Vals Finance Agent v2 — real legal and financial analysis, not synthetic tasks.
Long-context retrieval that holds up
On GraphWalks, which tests whether a model can actually use a 256K to 1M context, Google reports a lead of more than ten points over other flagships.
Long video understanding
LVBench, which measures comprehension over long video, is reported at 91.7% — the highest score Google lists for Argon.
Published numbers
Where Argon sits on Google's own benchmarks
These are vendor-reported results. Google has not released a thinking or effort setting for the headline scores, and no independent lab has replicated them yet.
Benchmark
What it measures
Argon
Nearest comparison
DeepSWE v1.1
Long-horizon software engineering
77.9%
Opus 5.5 at 74.2%, GPT-6 Astra at 74.1%
CWE-bench v1
Finding and patching vulnerabilities
68%
Ties for first place
AutomationBench
End-to-end business tasks, from Zapier
51.3%
Ranked first
LVBench
Understanding long video
91.7%
Highest score Google reports for Argon
Treat these as directional. Vendor benchmarks rarely survive contact with production workloads unchanged, and Bloomberg has reported that some Google employees are themselves sceptical that the benchmark lead transfers to real work.
On Felo
What a research workflow looks like with a 1M-token writer
Felo pairs search with document understanding, so a long-output model changes what a single session can produce.
Research that ends in a document
Gather sources, then have the model write the synthesis in one pass instead of assembling it section by section across many prompts.
Whole document sets in one task
Long-context retrieval is the capability Argon is strongest at, which is what makes contract sets, filings, and long reports tractable.
Drafting where tone matters
Argon's reported profile — strong text, moderate coding — suits teams that need prose that reads as written, not assembled.
Cost
What Argon costs, and what changes later
Google has published an introductory rate and a higher list rate, without saying when the switch happens. Plan your budget on the higher number.
Token type
Introductory
Later
Input
$2.00 / M
$4.00 / M
Cached input
95% off
95% off
Output
$10.00 / M
$20.00 / M
For comparison, the introductory rate sits roughly level with GPT-6.1 Sol and Sonnet 5.5, and near a fifth of GPT-6 Astra. Those are list prices, not measured cost per task — Argon answers harder questions, so a single run can still cost more.
How to work with Argon on Felo
01
Bring the whole question
Give Felo the document set or the long thread, not a summary of it. Argon's advantage starts where the context stops fitting in a prompt.
02
Ask for the finished artifact
With a 1M-token output ceiling you can request the full report or the complete draft in one instruction rather than in sections.
03
Check it against the sources
Felo keeps citations attached to the answer, which is what makes a long model-generated document reviewable.
Argon access in Felo Search is in progress. Until it opens, every other model in the Felo lineup is available now.
Open Felo AI SearchGemini 4 Argon: common questions
Not generally. Google released Argon first to selected partners in its Fairwind Program, and says wider access will start with paid API customers and Google AI Ultra subscribers. No date has been announced for that stage, and a paid subscription does not grant access today.
Try Google's frontier models on Felo
Search, read, and write with the model that fits the task. Argon joins the lineup when its rollout reaches Felo.
Open Felo AI SearchNo credit card required to use Felo AI Search.