Gemini 4 Argon is out at $2/$10 - until 30 November, then the price doubles

Google made Gemini 4 Argon generally available on 13 October at $2 per million input tokens and $10 per million output tokens. That price ends on 30 November. From 1 December it doubles to $4 and $20, according to Google's pricing as reported by VentureBeat and ComputingForGeeks.

If you are deciding whether to build on Argon this month, the second number is the one to plan around.

What happened

Argon was announced on 30 September for a small group of approved cybersecurity defenders. Two weeks later Google opened it on the paid Gemini API, Google AI Studio, Vertex AI, its Antigravity agent environment and eligible Gemini Enterprise plans. Google AI Pro and Ultra subscribers get it in the Gemini app. There is no free API tier. The model ID is `gemini-4-argon`, and it is also on OpenRouter.

The context window is 1,048,576 tokens, shared across input, thinking and output. ComputingForGeeks found one rate across the whole window, with no long-context surcharge. Batch and Flex run at half the standard price. Cached input is 95% off, though caching is implicit only; the explicit caching API is not supported for this model.

What is genuinely new

The independent numbers are good but not dominant. Vals AI puts Argon first on its Vals Index at 68.9%, ahead of Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%). It also ranks first on Vals' Finance Agent v2 at 65.4%. Artificial Analysis scores it 53 on its Intelligence Index, tied with GPT-6 Astra and behind Opus 5.5 at 58.

Google's own model card has one number that matters for this kind of work. On Zapier's AutomationBench, which tests multi-step business automations across real apps, Google reports Argon at 51.3% against 42.5% for Opus 5.5 and 41.4% for GPT-6 Astra. Long-context retrieval is its other clear lead: 84.2% on GraphWalks at 256K to 1M tokens, against 66.8% for Opus 5.5.

It is not ahead everywhere. On Terminal-bench 4.0, Opus 5.5 scores 66.4% to Argon's 57.4%. Most of the headline figures are Google's own, and Vals measured a lower score on one terminal benchmark using its own harness.

Artificial Analysis also reports a 15% hallucination rate on its Omniscience test, against 51% for GPT-6 Astra. If that holds up, it is the most useful thing in the release for anyone pulling facts out of documents.

What it means for a business owner

The workflows that benefit are the ones Argon's numbers point at. That means reading a large body of documents and acting on it: contract review across a whole deal folder, reconciling a quarter of invoices against purchase orders, answering support tickets from a long policy manual. A million-token window at a flat rate means fewer chunking tricks, and those tricks are where retrieval pipelines usually break.

The cost picture needs care. At the introductory rate, Argon is cheap for what it does. From December it costs the same as Claude Opus 5.5, according to ComputingForGeeks. A pilot you price in October will cost twice as much to run in December, with nothing else changed. If you can tolerate a delay, Batch at $1/$5 now and $2/$10 later is the realistic tier for overnight document work.

Output pricing includes thinking tokens. Google's thinking levels are low, medium and high, and medium is the default. ComputingForGeeks found that only high produced visible reasoning on short tasks. At high it took 12 to 13 seconds to the first token. Fine for a batch job, slow for a chat window.

The honest caveat

In ComputingForGeeks' testing, all 12 Kubernetes manifests Argon wrote failed to run. Opus 5.5 and GPT-6.1 Sol passed every one. Argon fixed them once shown the error. That pattern matters more than any benchmark. An agent that is right on the second try needs a loop that shows it the error, and someone has to build and pay for that loop.

The same tester hit empty responses. At high thinking with a small output cap, Argon spent the whole budget thinking and returned nothing. It also silently ignores `temperature`, `top_p` and `top_k`. If your existing pipeline depends on those settings, swapping the model ID will not be a clean switch.

There is also release pace. Google is reportedly testing newer Gemini 4 variants internally. Whatever you build should let you change models without a rewrite.

Wait if your automation already works on another model and the bill is acceptable. A few points on a benchmark do not justify a migration that re-tests every edge case.

What to do about it

If you have a document-heavy workflow that has been stuck on context limits, run it through Argon on Batch before 30 November. Log every run, compare its outputs field by field against your current model, and price the result at the December rate, not the October one. If it still wins at $4/$20, you have a real case. If it only wins at $2/$10, you have a promotion.