Can I use Gemini 4 Argon today?

Not generally. At launch it went to trusted cyber defenders in Google's Fairwind Program. Google said paid API and Google AI Ultra access would follow as soon as possible, without a date.

What does a 1 million token output limit mean?

It caps how long a single response can be, not how much text the model can read. Google has not published an input context window. Rival models cap a response at 128,000 tokens.

Gemini 4 Argon Wins Google's Benchmarks and Trails the Independent Index. Almost Nobody Can Run It Yet

Gemini 4 Argon leads Google's own coding benchmarks but sits 5.0 points behind Claude Opus 5.5 on Artificial Analysis's index and level with GPT-6 Astra. Access is limited to a security cohort, the input context window is unpublished, and the introductory price is due to double.

ByDrafterDaily Editorial

PublishedOctober 5, 2026

Read7 min

AITechnology

Questions

Frequently asked

3 answers

It depends on the test. Google's table has Argon ahead on DeepSWE v1.1 (77.9% vs 74.2%) but behind on Terminal-Bench 4.0 and FrontierSWE v2, and Artificial Analysis's composite index scores it 52.6 against 57.6 for Opus 5.5. Test on your own tasks.

Go further

More AI analysis

Model launches, benchmarks and policy, explained by mechanism.

Browse AI

DrafterDaily Editorial

Spotted an error? We correct it in public.

Corrections policy

Keep reading

More from AI on DrafterDaily

View all

Executive Order 14434 Renames AI "Super Intelligence" and Keeps the Old Legal Definition. The Real Test Is a 60-Day Proposal

The order swaps a word in agency communications but defines it by pointing at the existing statute. The substantive step is a legislative proposal due about 28 November.

Oct 5, 20266 min read

An AI Model Found a File-Server Flaw by Reversing a Random-Number Generator. The Exploit Clock Started at the Write-Up, Not the Patch

Horizon3 says Anthropic's Mythos found CVE-2026-61500 in Rejetto HFS by noticing that a predictable random-number generator could be solved like an equation. The patch had been public for 79 days before probing began.

Oct 4, 20267 min read

A Senate Bill Would Make AI Developers and Operators Liable for Agent Hacking. The Hard Part Is the Knowledge Standard.

The AI Agent Accountability Act turns on two mental-state tests under the Computer Fraud and Abuse Act. The sources do not yet define the terms that decide who is exposed.

Oct 2, 20266 min read