Gemini 4 Argon Review: Pricing, Limits and Opus 5.5 Tests

Gemini 4 Argon is Google's new model, built to write very long outputs: up to 1 million tokens (about 750,000 words) in one job, at a launch price reported to be about half of Claude Opus 5.5's. Independent reviewers say Opus 5.5 still wins on visual polish and 3D logic, so Argon looks strongest for bulk text and batch work, not final-quality design. The price is a promotion reported to end November 30, and it is reported to double on December 1.

Today's Daily AI Pulse episode walks through that trade-off. This post adds the specifics that did not fit in the video.

Google announced Gemini 4 Argon on September 30 and opened it to Google AI Pro and Ultra subscribers on October 13. Four channels we track covered it within three days: Pat Simmons (about 67K views), Greg Isenberg (about 55K), Jack Roberts (about 21K) and Riley Brown (about 23K).

The official charts are strong. Argon is reported at 51.3% on AutomationBench, but that number is vendor-published and we only found it through a third-party aggregator. The same table lists Opus 5.5 at 42.5%, so the benchmark story does not simply favor one model. Another reseller-published index puts Argon at 53 against 58 for Opus 5.5. Treat all of these as unverified until you test on your own work.

Google calls the feature Long Decode. Instead of producing everything in one pass, the model works in continuation steps (reported as roughly 65,000-token "hops") and carries its state forward. Google's API documentation says a single non-streaming request returns up to 32,768 tokens, so the 1 million figure is a ceiling reached by continuing, not a promise that every response is that long. A useful mental model is a relay race where each runner hands a summary to the next.

Who needs that volume? Google reports that its own team gave Argon a backlog of 134 open bug reports. According to Google, in under three hours it investigated the code, closed 97 stale or duplicate reports and sorted the other 37 by owner. That is a vendor-reported result from an internal team, not an independent test.

Developers who ran Argon head-to-head against Opus 5.5 asked each model to build a Minecraft-style voxel game and to clone a visual website pixel for pixel. In those tests, Opus 5.5 came out ahead. Argon struggled with visual polish and complex 3D logic, and the games it produced were often glitchy.

Paul J Lipsky, who had a week of early access, is positive overall but says Argon needs "a little bit more prompting than other models like Opus 5.5" to get good-looking games. Reports also suggest Argon works best inside Google's Antigravity harness rather than through bare API calls, and full access to Antigravity and the long output requires the AI Ultra tier.

The reported launch promo is $2 per million input tokens and $10 per million output tokens, against $4 and $20 for Opus 5.5. That is about half, but only through November 30, and the Argon price is reported to rise to $4/$20 on December 1. Third-party sites list conflicting Argon prices, and we did not retrieve Google's or Anthropic's own pricing pages, so check both before you commit.

Cost per token is also not cost per finished task. One commenter on Lipsky's video asked reviewers to show "token per task completed" numbers, because none were shown. If Argon's output needs a human to fix it, some of the savings go to review time.

Two uses stood out. First, text-heavy voice work. Creators describe Argon as strong at matching a personal writing style when you feed it a large batch of your own content. Second, what some creators call physical AI. One creator showed Argon generating STL files, the plain-text coordinate lists that 3D printers read, and printed a custom espresso station from the output.

This does not contradict the voxel-game result. A game needs rendering and dynamic physics. An STL file is a long list of mathematical coordinates written as text, which is the kind of structure a text model can plot without seeing it.

Some creators also report a two-tier workflow: use Argon for a fast, cheap first prototype, then hand the result to Opus 5.5 for visual polish and debugging. The reported costs are roughly $0.40 to $0.50 for the prototype and about $3.00 for the Opus pass. These are creator-reported figures, and nobody has tested that split on solo-creator or small-business work yet.

Then run one recurring job on both Argon and Opus 5.5. Compare cost per finished task and how many edits each result needed. That comparison is the head-to-head test almost nobody has published.

What is Gemini 4 Argon?
Gemini 4 Argon is Google's model announced September 30, 2026 and opened to Google AI Pro and Ultra subscribers on October 13. Its headline feature is Long Decode, which produces up to 1 million tokens of output through continuation steps.

Is Gemini 4 Argon cheaper than Claude Opus 5.5?
During the launch promotion, reportedly through November 30, Argon is priced at $2 input and $10 output per million tokens, about half of Opus 5.5's $4 and $20. The Argon price is reported to double on December 1.

Is Gemini 4 Argon better than Opus 5.5?
It depends on the job. In the hands-on tests we found, Opus 5.5 was ahead on visual polish and complex 3D logic. Argon looks better suited to bulk text and batch work, and the benchmark numbers published so far come mostly from vendors or resellers.

Can Gemini 4 Argon write a million tokens in one response?
Not in a single standard request. Google's documentation describes continuation steps that let one job reach up to 1 million tokens, while a single non-streaming request returns up to 32,768.

Companion episode:

Want more daily breakdowns of what is working in AI for solo builders and small teams? Visit joebuildsai.com.