Google DeepMind announced Gemini 4 Argon on 30 September, but at first only vetted cyber defenders could use it. On 13 October it became generally available: in the Gemini API, Vertex AI, Google's business products and the Gemini app for Google AI Pro and Ultra subscribers. It is Google's strongest model by a distance, with a few surprising strengths and two clear weaknesses. Here is what you need to know.
Where Argon leads, and where it doesn't
Google published results on 19 benchmarks against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Argon led or tied on 14 and trailed on 5. These are a few of the most relevant for business and data work:
| Benchmark | Argon | GPT-6 Astra | Opus 5.5 |
|---|---|---|---|
| AutomationBench (business workflows) | 51.3% | 41.4% | 42.5% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% |
| GraphWalks, 256K to 1M tokens | 84.2% | 71.8% | 66.8% |
| Chartography (reading charts) | 71.6% | 71.0% | 66.3% |
| Terminal-Bench 4.0 | 57.4% | 58.2% | 66.4% |
| FrontierSWE v2 (software engineering) | 55.0% | 65.5% | 62.3% |
The pattern is clear: Argon is strongest at long, multi-step business work over large amounts of text, and weaker at hands-on coding in a terminal. Note that these are Google's published results; on Artificial Analysis's independent index Argon scores 53, level with GPT-6 Astra and behind Claude Opus 5.5 (58) and Sonnet 5.5 (56).
The hallucination story
The most unusual result is on AA-Omniscience, an independent test of factual knowledge. Argon gave a wrong answer far less often than its rivals:
Source: Artificial Analysis. The trade-off: Argon answered only about 50% of questions correctly, against 63% for GPT-6 Astra, because it more often says it doesn't know.
For data and reporting work this is a meaningful difference. A model that says "I don't know" is easier to build on than one that confidently invents a figure.
How Google rolled it out
Pre-training starts
Google confirmed Gemini 4 training had begun.
Announced, cyber defenders only
Released through the Fairwind Program to governments, healthcare providers, telecoms and other priority defenders, without cyber guardrails.
Generally availableNow
Gemini API paid tier, Vertex AI, Google Antigravity, and the Gemini app for AI Pro and Ultra subscribers.
Introductory price endsUpcoming
API price rises from $2 / $10 to $4 / $20 per million tokens.
The 1M-token output limit, and why it matters
Most models can read a huge prompt but write only a few tens of thousands of tokens back. Argon can write up to a million, and Google added a "Long Decode Continuation" protocol so very long answers can be paused and resumed across API calls. In practice that opens up jobs such as:
One catch: input and output share the same 1M-token window, so a very long prompt leaves less room for the answer.
Price: what the change on 30 November means
Introductory (to 30 Nov)Standard
At the introductory price Argon is one of the best-value frontier models available. At the standard price it costs the same per token as Claude Opus 5.5, which scores higher overall. If you are evaluating it, now is the time.
Should you use it?
Good fit
Use Argon- Long documents, contracts and reports
- Multi-step business and finance workflows
- Reading charts and extracting data
- When avoiding confident wrong answers matters
Less ideal
Consider others- Heavy coding agents in a terminal
- Software-engineering benchmarks led by Claude and GPT
- Tight budgets after 30 November
- Workloads that need a free tier
Check your understanding
1/3When did Gemini 4 Argon become generally available?
It was announced on 30 September for cyber defenders only, then opened up on 13 October.
2/3What is unusual about Argon's output limit?
Earlier Gemini models were limited to 64K output tokens.
3/3Why is Argon's hallucination rate so low?
It answered about 50% of questions correctly but rarely invented an answer, scoring a 15% hallucination rate.
You got 0 out of 3.
Compare Argon with every other frontier model in The best AI models in October 2026, and see OpenAI's response in GPT-6.1 Sol and DevDay 2026.
Sources: Google DeepMind's announcement (30 September 2026) as summarised by eesel AI; MarkTechPost on general availability (13 October 2026); Artificial Analysis Intelligence Index and AA-Omniscience. Vendor benchmarks are Google's own results.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 14 October 2026. About me