Gemini 4 Argon: Google's frontier model, priced for cached agent loops
Date: 2026-10-14 · Source: Google DeepMind model card + Google launch post + Google Developers case study + AI Breakfast + Simon Willison · Raw: labs 10-13 23:02, labs 10-14 12:29, Gmail newsletters (private)
TL;DR
Gemini 4 Argon is Google DeepMind's new frontier model, generally available on 2026-10-13 after a cyber-defenders-only preview (the Fairwind Program) since 09-30. The model card describes a large sparse mixture-of-experts transformer (MoE: each token runs through a small subset of expert sub-networks), natively multimodal, trained from scratch on TPUs with JAX and Pathways, with a combined 1M-token input+output window and a March 2026 knowledge cutoff. Google's own table puts it first on the Vals Index (68.9% vs 67.0% Opus 5.5) and ahead on DeepSWE v1.1 (77.9% vs 74.2%), long-context GraphWalks 256K-1M (84.2% vs 66.8%) and Harvey's legal agent benchmark (19.6% vs 3.8%), but behind Opus 5.5 on Terminal-Bench 4.0 (57.4% vs 66.4%) and behind GPT-6 Astra on FrontierSWE v2 (55.0% vs 65.5%). The price is the story for this wiki: $2/$10 per million input/output tokens until 2026-11-30, cached input 95% off ($0.10/M), then $4/$20 from 2026-12-01.
flowchart LR
P["Agent turn<br/><small>repo, tools, history</small>"] --> C{"Prefix cached?<br/><small>same prefix as last turn</small>"}
C -->|yes| H["Cached input<br/><small>$0.10 per M</small>"]
C -->|no| F["Fresh input<br/><small>$2 per M, $4 Dec 1</small>"]
H --> M["Gemini 4 Argon<br/><small>sparse MoE, 1M window</small>"]
F --> M
M --> O["Output<br/><small>$10 per M, $20 Dec 1</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
class P input
class C loop
class H exit
class F err
class M core
class O exit
linkStyle 1 stroke:#2f9e44,stroke-width:2px
linkStyle 2 stroke:#e03131,stroke-width:2px
Key facts
- Availability: Gemini API, AI Studio, Antigravity (Ultra), Gemini app and AI Mode (Pro/Ultra), GitHub Copilot GA, Gemini Enterprise. Fairwind cyber-defender program expands to hundreds more partners.
- Model card benchmarks (Argon / GPT-6 Astra / Fable 5.1 / Opus 5.5): Vals Index 68.9 / 63.1 / 65.8 / 67.0; AutomationBench 51.3 / 41.4 / 31.4 / 42.5; DeepSWE v1.1 77.9 / 74.1 / 67.4 / 74.2; FrontierSWE v2 55.0 / 65.5 / 56.3 / 62.3; Terminal-Bench 4.0 57.4 / 58.2 / 57.9 / 66.4; PostTrainBench 45.3 / 44.3 / 40.2 / 49.3; GraphWalks 256K-1M 84.2 / 71.8 / 65.0 / 66.8; CWE-bench 68.0 / 68.0 / 58.0 / 67.0.
- Independent reads split: Vals puts it first; The Decoder says it trails Opus 5.5 and Meta's Muse Wonder Full on independent indices; Agent Arena places it 5th.
- Internal use claims (self-reported): 100K+ Google employees monthly; agentic code submissions +35%, rollbacks down about a third over four weeks; a Chromium refactor of a
6M-line legacy folder; Gemma 4 E2B on A100 via vLLM sped up 2.85x end to end (37,900 to 108,000+ tok/s) with token parity, as upstream PRs. - ADK overnight run (the honest case study): 11 unattended hours on google/adk-python (205K lines, ~800 files, ~17K tests). Six optimizations proposed, two shipped (in-memory session prep ~40% faster in long sessions, cold imports 21% faster), four rejected in human review.
- Safety: the model card says the full Safety & Security report follows separately; testing under the Frontier Safety Framework, human and automated red teaming. AI Breakfast reports a 0.7% attack success rate on Gray Swan's indirect prompt-injection suite over 15 adaptive attempts.
How it relates to prior wiki pages
- Confirms the cache-pricing thread. The 10-13 IBM agent-factory result (below, via compute economics) and this launch point the same way: for agent loops, the cached-input price and the prefix hit rate matter more than the list price. A 95% cache discount is the steepest the wiki has recorded from a frontier lab.
- Extends the frontier-crowding pattern from Meta Muse Wonder (10-13): two frontier launches in two days, each winning on its own chosen benchmarks and losing Terminal-Bench to Opus 5.5.
- Model card is architecture-light, like Muse Wonder's: "large sparse MoE" with no expert count, active parameters or attention design.
Open questions
- Does the 95% cache discount survive the 12-01 price doubling, and what hit rate do real agent harnesses get on it?
- The 2.85x Gemma speedup PRs are public. Do they hold on H100/B200, or are they A100-specific tuning?