Skip to content
FrontierDot

Google Argon Timeline: What Is Verified About Gemini 4 Argon

Live timeline · last updated · 1 entry in the update log

If you have seen “Google Argon” in headlines, the product is Gemini 4 Argon (API id gemini-4-argon). This post separates what Google has written on its own pages and official X accounts from what has only been said by others.

Google announced Gemini 4 Argon on Sep 30, 2026 to trusted cyber defenders and made it generally available on Oct 13, 2026 through the Gemini API, Google AI Studio, Antigravity, the Gemini app and AI Mode, at an introductory $2 / $10 per million tokens until Nov 30; the benchmark claims are Google’s own except where a named third party published them.

Read on 2026-10-18: Google’s two Gemini 4 Argon blog posts, the Gemini API model page and “What’s new” page (last updated 2026-10-16 UTC), the Google DeepMind model card, and official X posts by @Google, @GeminiApp, @GoogleDeepMind, @googledevs, @antigravity and @arcprize. Times are UTC as X reports them. We ran no benchmark ourselves and cite no third-party articles for any fact in the timeline.

Timeline

  1. Google announces Gemini 4 Argon in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect. It calls Argon a new frontier model, “not a modification or a fine-tune of a prior model” according to the model card, and says it is rolling out first to a set of trusted cyber defenders through the Fairwind Program. The post says Google is engaged in the U.S. government’s voluntary process for pre-release model access, sets the output limit at 1M tokens up from 64K, and states an introductory price of $2 per million input tokens and $10 per million output tokens.

    Google blog: Gemini 4 Argon, our next era of frontier intelligence

  2. The same post publishes Google’s first benchmark claims: 77.9% on DeepSWE v1.1 (long-horizon software engineering), first place on the Vals Index, 51.3% on Zapier’s AutomationBench, 91.7% on LVBench for long video understanding, and a tie for first at 68% on CWE-bench v1 for fixing security vulnerabilities. It also describes internal use at Google, including C/C++ to Rust migration work and a memory-savings project that it says freed more than 300 TiB once rolled out.

    Google blog: Gemini 4 Argon, our next era of frontier intelligence

  3. Google publishes “Gemini 4 Argon is now available for everyone”, dated Oct 13, 2026. It says Argon is globally available through the API in AI Studio and the Gemini platform, to Google AI Ultra subscribers in Google Antigravity, and to Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. It adds that expansion to hundreds more Fairwind partners begins that day and that 10,000 scientists can apply for free top-tier Antigravity usage for six months. The @googledevs account posts a developer walkthrough at 19:13 UTC.

    Google blog: Gemini 4 Argon is now available for everyone

  4. Google’s official X account runs a thread of Argon demonstrations starting around 14:55 UTC. One post says Argon was used to speed up the open Gemma 4 E2B model 2.85 times on an A100 through vLLM, from about 37,900 to over 108,000 tokens per second, with identical outputs. Later posts show an animated storybook built from an academic PDF (15:15 UTC), drug-discovery document synthesis at Isomorphic Labs (16:15 UTC) and a slide deck built from 80,000 NASA archive files (17:54 UTC).

    @Google on X: Argon speeds up Gemma open models

  5. ARC Prize publishes verified results for Gemini 4 Argon at 16:22 UTC: ARC-AGI-3 at 97.5% with a provider adapter harness (cost $4.8K) and 49.8% with the standard harness ($5.8K), ARC-AGI-2 at 91.3% and ARC-AGI-1 at 98.5%. ARC Prize is the benchmark’s own maintainer, so this is a primary source for the score, not for Google’s claims about the model.

    @arcprize on X: Gemini 4 Argon on ARC-AGI (Verified)

  6. At 22:52 UTC the official @antigravity account says Gemini 4 Argon is rolling out in Antigravity to paid Google AI Pro subscribers, with a separate weekly-only Argon quota on top of the existing Flash bucket, and says demand had been high. Antigravity’s Varun Mohan posts the same news at 22:52 UTC.

    @antigravity on X: Argon rolling out to Pro subscribers

  7. Google DeepMind posts at 16:11 UTC that four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours, working in a sandbox with only code and text, writing 2,783 programs, dividing jobs over chat and debugging together with no human help. The post quotes an article by one of the researchers. The Gemini app account separately shows Canvas turning dense PDFs into interactive web apps with Argon at 17:29 UTC.

    @GoogleDeepMind on X: four Argon agents beat Minecraft

  8. The Gemini API “What’s new” page is last updated on Oct 16 UTC. It lists the model id gemini-4-argon, a unified 1,048,576-token window, long decode up to about 1M tokens through background execution, thinking levels low, medium (default) and high, and says introductory pricing runs through Nov 30, 2026 before moving to $4 and $20 per million tokens from Dec 1. It also states that Argon is not on the Free tier.

    Gemini API docs: What’s new in Gemini 4 Argon

Background: which “Argon” is this?

Several unrelated things are called Argon, so the first job is to say which one the sources describe. The one Google has published is Gemini 4 Argon, the first named model in the Gemini 4 family. Google’s own model card describes Gemini 4 as “the next generation in the Gemini series, a suite of highly-capable, natively multimodal, reasoning models”, and describes Argon as a large sparse mixture-of-experts transformer with native multimodal support for text, vision and audio inputs. In the API it is called gemini-4-argon. Nothing we found at Google uses “Argon” for a separate product, a research codename that preceded this launch, or an earlier model.

The shorthand “Google Argon” is how the model gets referred to in conversation, but Google never calls it that. Its blog says “Gemini 4 Argon” in the title and then just “Argon” in running text, and the docs use the full name. If you are searching for primary sources, use the full name and the model id; the short form pulls in unrelated results.

Google positions Argon as its flagship reasoning model for real-world software engineering, enterprise knowledge work such as legal and finance, scientific work and cybersecurity defense. The model page lists text, image, video, audio and PDF as inputs and text as the only output, and states that image generation, audio generation and the Live API are not supported. Argon reads video and audio, but when Google’s demos show images, audio or 3D scenes, those come from other models (the blog names Nano Banana for images, Lyria and Gemini TTS for audio) driven by Argon as an orchestrator inside tools such as Antigravity.

The sequence: a phased launch

The launch ran in two clear phases that Google itself separated. On September 30 the blog post framed Argon as “rolling out to a set of trusted cyber defenders through our Fairwind Program”, and said that safely releasing capabilities at this level “requires a phased approach”. It also states that Google was “actively engaged in the U.S. government’s voluntary process for pre-release model access while we gradually expand access.” The post is explicit that developers, enterprises and consumers would get the model later, starting with paid API customers and Google AI Ultra subscribers.

On October 13 the second post, titled “Gemini 4 Argon is now available for everyone”, moved the model to general availability. It lists the channels: the API in AI Studio and the Gemini platform for developers and enterprises, Google Antigravity for Google AI Ultra subscribers, and the Gemini app and AI Mode in Search for Google AI Ultra and Pro subscribers. For enterprises it adds Gemini for business. Two programs ride along: scientists can apply for the highest Antigravity usage tier free for six months (10,000 places, rolling onboarding, subject to eligibility and capacity), and cyber defenders can apply to the Fairwind Program for access without cyber guardrails.

After that the cadence is visible on X. Over October 15 and 16 Google’s accounts posted a stream of demonstrations, the ARC Prize organization published its own verified scores, and Antigravity extended Argon to paid Google AI Pro subscribers with a separate weekly quota. These were rollouts and showcases of capabilities already described on October 13, and no new model was announced. The dated nodes above give each one with its link.

Two different dates for the launch

Google’s “available for everyone” post carries the date October 13, 2026. The earliest @googledevs demonstration post we found is also from October 13 (19:13 UTC), but the main @Google thread that tells people to “learn more about Gemini 4 Argon” and links to the blog is dated October 15 (15:15 UTC), and @antigravity’s Pro rollout is October 15 at 22:52 UTC. In the Google reply thread of October 15, the account also repeats the opening of applications for researchers and cyber defenders. We cannot tell from public pages whether the blog date reflects a staged publication or consumer-facing messaging that came two days later. We therefore date general availability to October 13 on the strength of the blog post and the docs, and we report the October 15 posts as what they are: the date those announcements appeared on X.

We keep the discrepancy visible here. If you rely on a specific date for a contract, a procurement document or a citation, quote the Google blog URL and its stated date, and check the Gemini API changelog for your own account’s availability, because eligibility varies by plan and region.

The numbers Google has published, and who stands behind each

Read the last column before repeating any figure. Most benchmark numbers come from Google’s own blog and are Google’s measurements, some on third-party benchmarks run by the benchmark’s owner and some on internal or partner evaluations. The ARC Prize figures are labelled verified by the maintainer and show a large gap between the provider-adapter harness result and the standard-harness result for ARC-AGI-3, so the same model scored very differently depending on the scaffold around it.

ClaimValueWho published it
Context window1,048,576 tokens, input and output unifiedGemini API model page
Max output65,536 default; up to 1,048,576 with long decodeGemini API model page
Intro price to Nov 30, 2026$2 input / $10 output per 1M tokensGemini API docs and Google blog
Price from Dec 1, 2026$4 input / $20 output per 1M tokensGemini API docs
Knowledge cutoffMarch 2026DeepMind model card
DeepSWE v1.177.9%Google blog (Sep 30)
AutomationBench51.3%, ranked #1Google blog (Sep 30), Zapier’s benchmark
LVBench91.7%Google blog (Sep 30)
CWE-bench v168%, tied for firstGoogle blog (Sep 30)
ARC-AGI-2 / ARC-AGI-191.3% / 98.5%ARC Prize on X (Verified)
ARC-AGI-397.5% with provider harness, 49.8% standard harnessARC Prize on X (Verified)

What Google says about safety

Google’s October 13 post describes a multi-layer defense stack: jailbreak resistance, monitoring and enforcement under its Frontier Safety Framework, a misalignment monitor that reads the agent’s reasoning and actions, and training meant to make the model report negative findings honestly and pause before irreversible steps. It reports that shortcut attempts on coding benchmarks fell to 2%, that external bio red teaming could unlock less than 1% of tested bio sub-tasks, and that attack success on Gray Swan’s indirect prompt-injection benchmark fell to 0.7% across 15 adaptive attempts. These are Google’s statements about its own testing. The model card adds that Argon may still hallucinate and may show occasional slowness or timeouts.

Google also says that for trusted defenders and its own teams it will release Argon without cyber guardrails. The public model is the guarded one. If you read a demonstration of unguarded vulnerability hunting, it is about the Fairwind configuration, not what a paid API customer gets.

What is not verified

  • Any release date for other Gemini 4 models. Google says only that Argon is “just the beginning” and that it is looking forward to “the rest of the Gemini 4 family”; no names, dates or specs are given.
  • Independent replication of the headline benchmark numbers other than the ARC Prize results. Google’s DeepSWE, AutomationBench, LVBench and CWE-bench numbers are published by Google in its blog; we did not find third-party leaderboards we could cite on primary pages.
  • Architecture details beyond “sparse mixture-of-experts transformer”: parameter count, expert count and training data are not disclosed on the model card or blog.
  • Claims by third parties that Argon is “insane at 3D”, “set a new SoTA” on a given product, or is available in a specific partner product. We saw such posts on X, including a repost by Google executives, but they are testimonials from others, and a repost is not Google verifying the claim.
  • Real-world impact figures from named partners (for example Wiz or Gamow Labs) beyond what Google’s post states; these are described in Google’s blog and we did not find the partners’ own pages.
  • Availability in a country, school account or enterprise tenant. Plans and rollouts differ, and the blog says only “globally available”.

How this fits with other model timelines on Front

Argon is one of several frontier or near-frontier releases we have tracked from primary sources this month. For the xAI side of the story, with a model that has been talked about but not released, see Grok 4.8: what xAI has actually said. For open-weights releases with real weights you can download, compare it with the Reflection Beam 501B timeline and the Meta Muse Spark Open timeline. For practical steps once you want to try Argon, read What you can do with Google Argon today, and for the way these labs are being used on open maths problems see the AI mathematicians timeline. Google’s Argon demos also use Blender through MCP, so our Blender alternatives guide is relevant if you want to pair the model with a 3D tool. The blog index lists everything.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

What is Google Argon?

It is Gemini 4 Argon, Google DeepMind’s flagship reasoning model, announced on September 30, 2026 and made generally available on October 13, 2026. Its API model id is gemini-4-argon.

Is Argon the same as Gemini 4?

No. Argon is the first model of the Gemini 4 family. Google says the rest of the family is coming but has not named other models or given dates.

How much does Gemini 4 Argon cost?

Through November 30, 2026 the API price is $2 per million input tokens and $10 per million output tokens, then $4 and $20 from December 1. It is not available on the free API tier.

Who can use it?

Per Google’s blog: developers and enterprises through the Gemini API and AI Studio (paid tier), Google AI Ultra subscribers in Antigravity and AI Studio, Google AI Ultra and Pro subscribers in the Gemini app and AI Mode, and Pro subscribers in Antigravity as announced by @antigravity on October 15.

Are the benchmark scores independent?

Mostly not. Google published most scores itself. ARC Prize published the ARC-AGI results as verified by the benchmark’s maintainer.

Update log

  • — First published. Checked Google’s blog, Gemini API docs, DeepMind model card and official X posts through Oct 18, 2026 (UTC+8).

Sources