Gemini 4 Argon first look: what it does better and when to switch

Google announced Gemini 4 Argon on September 30 and opened it to everyone on October 13, with a launch price that doubles on December 1. My competitor tracker logged Riley Brown's Argon video at 130,849 views on October 12, so the audience is already asking whether to move. I spent the window pulling numbers I could verify - benchmarks, prices, and two weeks of builder complaints - and sorted them into what changes for a one-person operation and what only fills thumbnails.

Published October 27, 2026 · By Dmitry Tolstoi · 9 min read

Two dates and one loud thread

In short: The rollout came in two steps, the community response is measurable, and the cheap-pricing window closes on November 30.

The release ran on two dates. On September 30 Google announced Gemini 4 Argon and handed it first to security teams in its Fairwind Program. On October 13 it became broadly available through the regular API. Between those dates the discussion got loud: the Hacker News launch thread collected 1,705 points and 1,199 comments, and my own tracker logged Riley Brown's video "ChatGPT Dots Is Insane... But Gemini's NEW Argon Is EVEN Bigger" at 130,849 views on October 12, plus AICodeKing's benchmark breakdown at 31,297 views a week earlier. That is the same demand pattern I saw before the ChatGPT Dots wave: viewers are asking a question, and most answers they get are thumbnails. Here is what I could document instead.

What shipped on October 13?

In short: Argon now sits in the standard Gemini API and most Google surfaces, with a unified limit of up to 1 million tokens.

Since October 13 the model answers as gemini-4-argon in the Gemini API, Google AI Studio, the Antigravity development environment, Android Studio, Firebase AI Logic, the Gemini app and eligible enterprise subscriptions. The specification that matters for builders is the token ceiling: Google describes a unified limit of up to 1 million input and output tokens, with a continuation protocol for jobs that outgrow a single response. The output side used to cap at 64,000 tokens, so this is the first Gemini that can hand you a very large artifact - a migration diff, a full report, a long transcript - in one run instead of twenty stitched ones.

What does Argon cost before and after November 30?

In short: Budget on the December price of $4 and $20 per million tokens, and watch output length - Argon spends tokens generously.

The launch price is a timer, and Google published the whole schedule upfront:

Token price is half the story. In eesel's test runs Argon averaged 62,000 output tokens per task where GPT-6 Astra spent 27,000, which put a typical task at $1.99 on launch pricing and around $3.98 after the increase. The model is verbose by default, so a per-token bargain can still lose a per-task comparison. If you test during the cheap window, log cost per finished task, and double it before deciding anything - December you will thank October you.

Where Argon actually wins

In short: Verified wins are long-document work, multi-step automation and a low rate of confident guessing - the profile of an operations model.

Three results stand out in independent testing rather than vendor slides. First, long documents: on GraphWalks retrieval across 256,000 to 1 million tokens of context, Argon scored 84.2% against 66.8% for Claude Opus 5.5 and 71.8% for GPT-6 Astra in eesel's compilation. If your pipeline digests contracts, codebases or transcript archives, that gap is the headline. Second, automation chains: on AutomationBench-AA, the Artificial Analysis version of the business-automation suite, Argon posted 77.5% against 71.3% for Claude Sonnet 5.5 and 69.5% for Opus 5.5. Third, honesty under uncertainty: Artificial Analysis measured a 15% hallucination rate against 51% for Astra, with a visible trade-off - Argon answered 50% of questions correctly against Astra's 63%, because it says "I don't know" where others guess. For an unattended pipeline I will take that trade every time; a wrong answer costs more than a refusal.

The aggregate indexes split. Vals AI ranks Argon first on its GDP-weighted index at 68.90%, just ahead of Claude Sonnet 5.5 at 67.04%. Google's own numbers - 77.9% on DeepSWE v1.1 for software engineering, 91.7% on LVBench for long video, a claim that internal teams raised agentic code submissions 35% while cutting rollbacks 30% - read well, and they are vendor numbers until someone reproduces them.

Where it loses on the same benchmarks

In short: Terminal-heavy agent work still belongs to Claude, and on the composite intelligence index Argon lands mid-pack, tied with Astra.

The same compilations show the ceiling. On Terminal-Bench 4.0, the shell-and-tools agent suite, Argon scored 57.1% - fourth place, behind Claude Sonnet 5.5 at 63.6%, Opus 5.5 at 59.6% and Astra at 59.1%. If your agents live in a terminal, the new model is a downgrade, and the numbers say so plainly. On the Artificial Analysis Intelligence Index, Argon sits at 53, tied with GPT-6 Astra and below both Opus 5.5 at 58 and Sonnet 5.5 at 56. Sam Witteveen, CEO of Red Dragon AI, put the practitioner consensus into one line in VentureBeat's coverage: "This is a really, really, really good model. But is it the best model out there? I actually don't think it is."

What builders hit in the first two weeks

In short: The loudest complaints target the tooling around the model - quota opacity, forced compaction and prompt friction - rather than the model itself.

The Hacker News thread is worth an hour of your reading time, because the pain is specific. Quota visibility: usage limits refresh roughly every 30 minutes and sit four clicks deep, so you learn you are out of budget by hitting the wall. Context handling: Antigravity forces a compaction at 250,000 tokens, which interrupts exactly the long-horizon jobs the model is sold on. Workflow friction, in commenter piyh's words: "Antigravity queues all my prompts and makes me confirm to submit them anytime a long running process like a hot reloading UI is active." And there are reports of account-level bans for routing API access through unofficial paths - one commenter warned the ban can cover the entire Google account. Several people in the thread tried moving their coding sessions over from Claude Code and moved back within days. The sharpest summary came from commenter mapontosevenths: "If they didn't force you to use a terrible harness, charge too much for way too little...Google could take over the AI market."

When switching makes sense

In short: Route individual workloads, keep the stack mixed: long-document and automation jobs go to Argon, terminal agents and architecture stay put.

Read the benchmarks as a routing table instead of a verdict. A job that reads hundreds of pages and must not invent what it did not find - contract review, research digestion, log analysis - is Argon's strongest verified case, and the 1 million token output means the result arrives whole. A chain that moves data between services on a schedule matches the automation scores. A coding agent that lives in a shell stays on Claude: the Terminal-Bench gap is real, and my own pipeline keeps its Opus 5.5 setup for architecture work for exactly that reason. Writing and frontend sit in taste territory; thread commenters rate Gemini well there, and no benchmark settles it for you. That mirrors how experienced builders in the thread actually operate - several run two or three subscriptions and route by task type. One hygiene rule before any test: use a separate Google account and the official API, because the ban reports above are not the place to discover your mail depends on an experiment.

A cheap test before December 1

In short: One workflow, one week, a hard budget cap, numbers logged against your current model - all inside the launch-price window.

The intro price makes October the cheapest month to find out whether any of this applies to your operation. The test I would run, and the one I am scheduling for my own document pipeline:

1. Pick one workflow with a measurable output - a weekly report, a batch of support answers, a document digest. One, and only one.
2. Create a fresh Google account for it and set a billing cap, so the experiment cannot touch your main account or your month's budget.
3. Run the same inputs through your current model and through gemini-4-argon for one week, with cached input enabled where the job repeats context.
4. Log cost per finished task, output token counts and error types - a guess that went wrong counts double, a refusal counts half.
5. Double Argon's cost column for December pricing, then decide. If the sheet does not show a clear win, the switch is marketing, and you skip it.

Whatever the sheet says, you end the week owning a number instead of an opinion, and the whole exercise runs a few dollars at launch prices.

Where this comes from

Release dates, platform list and pricing are from Google's announcement and the Google Cloud launch post, both read on October 27, 2026. Independent benchmark figures - GraphWalks, AutomationBench-AA, Terminal-Bench 4.0, the hallucination rates and the Intelligence Index - are from eesel's review and Artificial Analysis data as reported by VentureBeat and Techmeme. The Vals AI ranking and the Sam Witteveen quote are from VentureBeat. Builder complaints and both commenter quotes are from the Hacker News launch thread, which stood at 1,705 points and 1,199 comments when I read it. View counts are my own tracker's log for October 5 and October 12, 2026. I have not moved my own pipeline to Argon yet; when the document-digest test above runs, the before-and-after lands here the same way. Nothing above is a promise about how your workload will behave.