Google announced Gemini 4 Argon on September 30, 2026, the first model of the Gemini 4 generation and, by Google's account, its most capable to date. It is built for long, multi-step work: software engineering, legal and financial knowledge work, and finding and patching security vulnerabilities. You cannot call it yet unless you are a vetted cyber defender.
The short version
| Gemini 4 Argon | |
|---|---|
| Announced | September 30, 2026, by Koray Kavukcuoglu, Google DeepMind |
| Output limit | 1 million tokens, up from 64K |
| Introductory price | $2 per million input tokens, $10 per million output |
| Standard price | $4 input, $20 output, after the introductory period |
| Cached input | 95% off the input price |
| Access today | Trusted cyber defenders, through Google's Fairwind Program |
| Next | Paid API customers and Google AI Ultra subscribers; no date given |
Google has not published a model ID or an end date for the introductory price.
Where it leads, and where it does not
Google published 19 benchmarks against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1. Argon leads 13 of them. The wins cluster in long-context work, business automation and knowledge work. On terminal-driven agent loops and the hardest software engineering set, it trails.
Gemini 4 Argon against three frontier models, on five of the 19 benchmarks Google published
DeepSWE v1.1
- Gemini 4 ArgonLead77.9
- GPT-6 Astra74.1
- Claude Opus 5.574.2
- Claude Fable 5.167.4
AutomationBench
- Gemini 4 ArgonLead51.3
- GPT-6 Astra41.4
- Claude Opus 5.542.5
- Claude Fable 5.131.4
GraphWalks, 256K to 1M tokens
- Gemini 4 ArgonLead84.2
- GPT-6 Astra71.8
- Claude Opus 5.566.8
- Claude Fable 5.165.0
Terminal-bench 4.0
- Gemini 4 Argon57.4
- GPT-6 Astra58.2
- Claude Opus 5.5Lead66.4
- Claude Fable 5.157.9
FrontierSWE v2
- Gemini 4 Argon55.0
- GPT-6 AstraLead65.5
- Claude Opus 5.562.3
- Claude Fable 5.156.3
Two of those five go against Argon, and we kept them in on purpose. A model that leads a table still loses specific workloads, and an agent only ever runs specific workloads.
What changes for agents
Long outputs in one pass. A 1 million token output limit removes the stitching that large refactors and migrations need today, where an agent emits a change in pieces and re-prompts between them. Google says it has used Argon internally on C and C++ to Rust migrations of more than 800,000 lines.
Price at the frontier. The introductory rate puts a frontier model at $2 / $10. Per-token price is not cost per task, though: Google has not published per-task figures, and a model that reasons longer can cost more to finish the same job.
Prompt injection. Google calls Argon its most resilient model against indirect prompt injection and says it leads Gray Swan's benchmark for it. That lowers one risk for agents that read untrusted content. It does not replace least-privilege tool access, approval gates or sandboxing.
What to do before you route traffic to it
We have not run Argon; outside the Fairwind Program, nobody has. When access opens, treat it like any other model change: a release that has to pass your own evaluation before it carries production work.
A new model is a release. It goes through evaluation before it goes near production.
- MODEL.NEWA model shipsVendor benchmarks, price and limits. Useful, but not your workload.
- EVAL.RUNRun your own suiteYour golden dataset, your tools: task success, tool-call accuracy, cost per task, latency.
- MODEL.ROUTERoute by workloadSend it the tasks it wins on your numbers. Keep the rest where they are.
- STATUS: ONLINERelease with rollbackVersioned, traced, and reversible if behaviour regresses.
In practice that means three things. Keep an evaluation suite that can score a new model on your tasks in hours, not weeks. Put models behind a routing layer so that switching one is a configuration change. And keep agent security controls independent of whichever model is behind them.