Hauke Hillebrandt

Oct 2026

What Twitter says about Gemini 4 Argon

Open in Google Docs

Hauke Hillebrandt, 18 October 2026. Research done with Claude, using Scry's Twitter archive.

We read what people on Twitter (X) said about Google's Gemini 4 Argon from September to 18 Oct 2026. The posts come from Scry's Twitter archive, pulled with free queries on 18 Oct at ~11:30 Coordinated Universal Time (UTC). We also looked for posts on mathematics, recursive self-improvement (RSI) and the model card.

What we read

Findings

  1. Two spikes, both set by Google. Google announced Argon on 30 Sep for a set of trusted testers and released it worldwide on 13 Oct. The two-day launch windows hold ~2/3 of all cohort tweets. Posts by Google and its executives drew about a third of the likes on the 120 most-liked posts, led by Sundar Pichai's (~28k).
  2. Jokes led the first spike. The most-liked post not from Google is a Japanese joke (8,563 likes): Argon seems so self-critical that it was probably abused in training.
  3. More praise than complaint, mostly news. In the random sample, about a third of on-topic posts are positive and a sixth negative (63 and 32 of 198 in one coding). Most of the rest is news without a verdict.
  4. Benchmarks put it near the top. Epoch's Capabilities Index gives Argon 165, behind Opus 5.5 (167) and GPT-6 Astra (166). The Abstraction and Reasoning Corpus (ARC) Prize verified 97.5% on ARC-AGI-3 with the provider's adapter, but 49.8% with the standard setup.
  5. Testers are split. Ethan Mollick calls it "the first non-OpenAI/Anthropic model to pass the Fable/Mythos/Astra bar". Wes Bos rates it good but "not Opus5.5/Astra/Fable levels". Theo calls it "a huge jump for Gemini" but, a day later, "outright unusable for real code work". One tester's bug-fixing test scored Argon 33 at launch and 40.7 four days later, against 41.7 for Opus 5.5.
  6. Cheap per token, wordy per task. Partners list $2 per million input tokens and $10 per million output tokens. Several posts say this introductory price doubles in December. One-prompt demos cost a fraction of Opus 5.5's, like $0.19 against $12.06 for a ship in a bottle. That tester still preferred Opus's result. And in one tester's build, Argon used ~50% more tokens than Opus 5.5.
  7. Access and quotas drew nearly half the complaints. Pro subscribers waited two days after launch for Antigravity, Google's coding tool, and some then reported a weekly quota that "can disappear in a day". Logan Kilpatrick (Google) said Argon's "big model smell" limits how widely Google can serve it. In the random sample, 14 of the 32 negative posts are about the wait, tier limits or quotas.
  8. RSI talk: a few Google voices, much hype. A Google research leader thanked the teams "Developing RSI loops and agents on top of this model" for memory optimization and code migration. A new Google DeepMind scientist announced joining "the RSI team". Sergey Brin's line that Google is "starting to get on that RSI treadmill" reaches the pull only through a news bot, after Kilpatrick posted a conversation with Brin. In September, three accounts posted the same line, "As you know Google Achieved RSI", and a self-described Gemini bull called it "made up". Only 1 of the 200 sampled posts mentions RSI, and it is an automated Grok reply.
  9. Google's speed-up figures are its own. Google says that across its codebases Argon brought a 35% increase in "agentic code" and 30% fewer rollbacks after submission. It also says Chrome engineers finished "months of work in just weeks" with it. A Bloomberg report, seen only through relays, said Gemini 4 "does less well when employees actually put it to work", and a Google DeepMind engineer reportedly called that "bs". No pulled post checks Google's figures.
  10. Mathematics is thin, and the hype outran the result. The main math result is a preprint in which Argon agents extended counterexamples to Grad's conjecture in plasma physics. Two papers had found the first ones the month before, one of them with GPT-6 Astra. A relay says the researchers chose the problem and "independently re-derived every mathematical derivation". A viral post instead claims that Argon "cracked a fusion energy math problem in a couple of hours". The one first-hand test of research math is negative: an Erdős attempt was "mostly incorrect". The same tester called a 78% score on FrontierMath Tier 4 not great, without naming its source. Elliot Glazer agreed by current frontier standards, noting how fast things have moved this year.
  11. Model-card talk is about safety. Two accounts that tested Argon report it writing "hello evaluators reading my chain-of-thought". An account that elsewhere calls it "our safety report" says Argon is "significantly less monitorable than past models". No pulled post mentions the model card's critical capability levels (CCLs) for machine-learning research and development (ML R&D). "CCL" matches none of the 6,723 cohort tweets.

Quotes

Dates are UTC, and likes are as of the pull.

Caveats