Gemini 4 Argon

TL;DR Highlight

Google unveiled Gemini 4 Argon, but opened it first only to trusted cyber defenders.

Who Should Read

Backend or agent developers using the Gemini API who want to know the new model's price, performance, and release timing before deciding whether to switch.

Core Mechanics

Output limit jumps from 64K to 1M tokens
- The output token limit rose from 64K to 1M tokens.
- Google says this lets the model think and generate hundreds of thousands of tokens to solve hard problems in one go.
Access starts with the Fairwind Program
- It is currently available only to trusted cyber defenders via the Fairwind Program.
- Developers, enterprises, and consumers come later, starting with paid API customers and Google AI Ultra subscribers.
- Google is also taking part in the U.S. government's voluntary pre-release access process.
Pricing: $2 input, $10 output
- The introductory price is $2 per million input tokens and $10 per million output tokens.
- Cached input tokens are 95% cheaper than the input price.
- After the introductory period, $4 input and $20 output apply.
77.9% on DeepSWE v1.1
- Google says it sets a new state of the art on DeepSWE v1.1, a real-world long-horizon software engineering benchmark, at 77.9%.

Community Reactions

Opinion that pricing is uncompetitive
- One commenter said Artificial Analysis shows average cost per task of $1.99, above Astra high ($1.73) and Opus 5.5 high ($1.82).
- They argued that even at discount pricing, weak caching on AI Studio and Vertex hurts competitiveness.
Info on low hallucination rate
- One commenter said Artificial Analysis shows Argon's hallucination rate at 15%.
- The same comment put recent OpenAI models in the 40s-50s and Anthropic in the 60s-70s.
Experience: good enough to hand off complex work
- A commenter with weeks of access said it was the first Gemini they could delegate complex, domain-research-heavy tasks to.
- They said it is not fully reliable but verifying its output is much cheaper than doing the work themselves.
Complaints about benchmarks and access
- Some said the benchmark results only show Max reasoning levels, which is far from real usage.
- Another subscriber complained they cannot use the latest models and do real work with Codex or Claude Code instead.

How to Apply

Move model names into config
- Today, move hardcoded model IDs into environment variables or a config file.
- Add a switch so the same input can run on two providers.
Simulate cost at Argon pricing
- Compute monthly cost at $2/$10 and at the $4/$20 regular price using your input and output tokens.
- Apply the 95% cached input discount and compare with your current model.
List long-output task candidates
- List tasks you currently truncate because of the 64K output limit, such as large code conversions or long drafts.
- Test them first when you get access.
Add injection tests to document-reading agents
- Build test cases that mix malicious instructions into documents your agent reads.
- Rerun the same cases as a regression test whenever you change models.

Try It

copy & run

# Prompt to paste into Claude or ChatGPT (simple Argon cost comparison)

Calculate costs from the monthly usage below and show a table.
- Monthly input tokens: {{input tokens}}
- Monthly output tokens: {{output tokens}}
- Cache-hit input ratio: {{0 to 1}}

Pricing:
- Gemini 4 Argon introductory: $2/1M input, $10/1M output
- After the introductory period: $4/1M input, $20/1M output
- Cached input tokens get 95% off the input price
- Current model: ${{current input price}}/1M input, ${{current output price}}/1M output

Compare introductory, regular, and current-model pricing and show the monthly difference.

Next: double the output tokens and recalculate.

Terminology

Fairwind ProgramGoogle's program giving trusted cyber defenders early access to Argon. Think of it as a pilot lane before general release.

Indirect prompt injectionAn attack hiding malicious instructions in web pages or documents to hijack a model. It is like slipping a fake instruction slip into a package.

chain-of-thoughtThe intermediate reasoning a model writes before its answer. It's like the scratch work in a solution.

SIMDA CPU feature that processes multiple data items with one instruction. Like running several conveyor belts at once.

Cached input tokensInput tokens reused from a previously sent prompt prefix at a lower price. Like keeping a copy of a manual you reuse.