Gemini 4 Argon
Released Sep 30, 2026 · knowledge cutoff unpublished
- Status
- Frontier
- Location
- United States / United Kingdom
- Type
- LLM
- Modality
- Multimodal
- Context window
- —
- Max output
- 1M
- Speed
- —
- Price ($/MTok in / out)
- $2 / $10
- Cost per task (high effort)
- $1.99
- Open weights
- No
Benchmarks1 of 10 reported
LMArena Elo
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Declines to answer obscure-fact questions far more often than it fabricates — it trades recall for honesty, so confident wrong answers are rare
- Can write a whole large artifact in a single response, with no continuation prompts where earlier Gemini models cut off mid-output
- Aimed at long-horizon agent work — Google already runs it internally on C/C++-to-Rust migrations spanning tens of thousands of lines
Weaknesses
- Thinks at great length — Artificial Analysis' runs average more than twice the output tokens per task of comparable frontier models, so the low per-token rate understates what a job costs
- Can't simply be picked up yet: access starts with vetted cyber defenders in the Fairwind Program, with paid-API and AI Ultra access promised but undated
- Behavior is a moving target this early — no model card, no public API model id and no hands-on reviews exist, so its voice, verbosity and agentic stamina are uncharacterised
Google's first Gemini 4 model, announced 2026-09-30 and rolling out first to trusted cyber defenders through the Fairwind Program while Google takes part in the U.S. government's voluntary pre-release testing; broader access is promised to paid API customers and Google AI Ultra subscribers 'as soon as possible', hence availability "restricted" and no apiIds (the Gemini API catalog lists no Argon model id yet). Pricing is the introductory $2/$10 per MTok (cached input 95% off), rising to $4/$20 afterwards; no end date for the introductory period was given. maxOutput is Google's stated '1M tokens' (up from 64K on earlier Gemini models) — the exact API token limit is unpublished; Google has not disclosed the input context window (Artificial Analysis lists 1M) or a knowledge cutoff, so both stay null. Google's evaluation PDF (deepmind.google/models/evals-methodology/gemini-4-argon) uses only untracked suites: Vals Index 68.9%, AutomationBench 51.3%, Vals Finance Agent v2 65.4%, Harvey Legal Agent 19.6%, DeepSWE v1.1 77.9%, FrontierSWE v2 55.0%, Vibe Code Bench 91.9%, Terminal-Bench 4.0 57.4% (not the tracked 2.1), PostTrainBench 45.3%, Terminal-Bench Science 0.1 57.6%, LABBench2 88.8%, RiemannBench 76.0%, GraphWalks 99.7% (to 128K) / 84.2% (256K–1M), Agent's Last Exam 39.5%, OSWorld 2.0 offline partial 69.2%, Chartography 71.6%, LVBench 91.7% and CWE-bench v1 68.0%; no GPQA Diamond, SWE-bench, HLE or ARC-AGI-2 figures. hle is null: Google's evaluation PDF has no HLE row, and Artificial Analysis' 57% (high) comes from its own 2,158-question text-only subset whose tool setting its methodology doesn't state, so it can't be pinned to a primary source. LMArena Elo 1525 (±9, preliminary, 4,942 votes) is the 'gemini-4-argon-high' text-arena listing, rank 1 on the 2026-09-30 update. Artificial Analysis scores it 53 on its Intelligence Index at high effort (the only variant AA lists) at $1.99 per task at introductory pricing ($3.98 at standard pricing), averaging about 62K output tokens per task against GPT-6 Astra's 27K, with an AA-Omniscience hallucination rate of 15%; AA publishes no speed figures and rates it 1611 on GDPval-AA v2.1, a different scale from the v2 this key records, so gdpvalAA stays null. predecessorId is null: Google names no model it replaces.
For developers
API model strings
Not researched
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
No recorded predecessor