Gemini 4 Argon Can Generate a Million Tokens in One Job

Ronni Holmvig Strøm · 2026-10-15

Google put Gemini 4 Argon in front of developers and businesses this week. Benchmark tables. Deep Sweep. A column of comparisons against Opus 5.5 and GPT 6.1. The number worth reading is in the API documentation instead: a single Argon response can run to 1,048,576 tokens, and input, thinking, and

The Unit of Work Is a Job, Not a Turn

Long decode is the feature, and its mechanics reward a close read. A single model call still generates at most 64K tokens. When a response needs more, Google's servers run it in 64K hops, each one resuming where the last stopped, each one appended as another model_output step under one interaction ID. You join the steps yourself to read the answer.

It works only in background mode. Set background: true, raise maxoutputtokens above 65,536, then poll. A synchronous call that asks for more returns a 400, and the docs say long decode for the streaming path is "coming soon," which is a polite way of saying the frontier path is asynchronous for now.

What Google shipped is a job runner wearing a model's API. Nobody watches these tokens arrive. One early run reported by MindStudio produced roughly a million tokens across 16 segments over more than three hours. Three hours is not a conversation. It is a build.

Input and Output Now Compete for the Same Budget

Previous frontier models gave you a million tokens of input and a separate 64K of output, which made context nearly free at the margin. Teams stuffed in the repository and the quarter's tickets. The ceiling on the answer never moved either way.

Argon prices that habit. The documentation is blunt about it: your output limit is maxoutputtokens, up to 1,048,576 minus your input. A 400K-token repository dump now costs you 400K tokens of code the model will not write.

Cost tightens the same screw. Explicit caching is not supported on Argon, only implicit, which takes away the lever teams use to make long-context agents affordable. Introductory pricing is $2 per million input tokens and $10 per million output, and Google's own page says both double on December 1. Thinking tokens bill at the output rate and draw from the same budget, so a high-reasoning long-decode run spends part of its million on text the user never sees.

Capability Arrived Before the Disposition

Early testers could not get near the ceiling by asking normally. Argon wraps up early, the instinct every assistant model has been trained into, and one tester watched it ration itself into fragments when left alone. Reaching a million tokens meant telling the model in the system prompt that it had a million tokens and should not summarize or stop short.

The capability lives in the weights and the serving stack. The willingness to use it does not. We spent three years teaching these systems that a good answer is a short one, rewarding concision at every stage of preference training, and the result is a model that will not spend its budget unless told the budget exists. Google's own docs warn that an agent framework left at a default of 8,192 will stop the model "before it finishes thinking."