Google DeepMind introduced Gemini 4 Argon on 30 September 2026. The headline change is not on the input side but the output side: the generation limit rises from 64,000 to one million tokens. The model goes first to cyber defenders rather than to all developers at once.
Key takeaways
- Launched 30 September 2026, model name Gemini 4 Argon
- A one million token output limit replaces the previous 64,000
- Introductory pricing: $2 per million input tokens and $10 per million output tokens
- After the introductory period, rates rise to $4 and $20 per million tokens
- Distribution starts with the Fairwind Program for trusted defenders, reaching developers and enterprises later
A million tokens, on the way out
Most models advertise a large context window, meaning how much can be read in. Argon moves a different slider: how much the model can generate in one pass. The jump from 64,000 to one million tokens targets long reasoning chains and tasks where the answer is itself a large artefact, a repository refactor for instance. Cached input carries a 95% discount against the base price.
Argon pricing, per million tokens:
What the benchmarks say
Google leads with professional work rather than knowledge tests.
| Benchmark | What it measures | Argon's score |
|---|---|---|
| DeepSWE v1.1 | software engineering tasks built from real issues | 77.9% |
| AutomationBench | automation of professional tasks | 51.3% — first place |
| LVBench | long video understanding | 91.7% |
| CWE-bench v1 | remediation of known vulnerability classes | 68% — tied with the best rival |
The model page on Google DeepMind sets Argon against GPT-6 Astra and Claude models.
Defenders first, everyone else later
The release mode itself is unusual. The model goes first to Fairwind Program?Fairwind: A Google programme giving trusted cyber defenders access to the model ahead of the broad release. participants, and authorised defenders can optionally have offensive cyber guardrails removed. Google also points to prompt injection robustness measured on the Gray Swan IPI?Gray Swan IPI: A test of resistance to indirect prompt injection — instructions hidden inside documents and pages the model reads while working. benchmark.
Why it matters
Shifting the emphasis from input to output changes what the model is practically good for. If a single response can be the size of a mid-sized project, there is less need to split work across dozens of calls and stitch results together outside the model. The cost of a single error rises, because verifying a million output tokens is harder than checking a short answer. Selective distribution shows vendors treating cyber capability as its own risk category.
What's next?
- The introductory price is temporary and rates double to $4 and $20 per million tokens once it ends, according to Google's price list
- Broader availability for developers, enterprises and consumers has been announced, but without a date
Sources
- Google — Gemini 4 Argon: our next era of frontier intelligence
- Google DeepMind — Gemini models
