# Gemini 4 Argon announced

AI & tech · Heat 1.7 / 10 · 9 shows · 1h 22m airtime

Google announced Gemini 4 Argon on September 28, 2026, its first new frontier model in more than six months, initially limited to trusted cyber defenders and enterprise cloud customers pending government safety evaluations, citing a 68% CWE Bench score. Bloomberg then reported insiders say it struggles with day-to-day tasks, and the restricted rollout drew a negative stock reaction. Google said it would be inaccurate to call Gemini 4 underperforming in coding, and Artificial Analysis puts Argon High's hallucination rate at 15% versus Astra's 51%. Argon agents migrated more than 800,000 lines in the Fuchsia OS Zircon kernel, and the model is now rolling out to the Gemini app and AI mode in search for Google AI Pro subscribers.

Page: https://boilpoints.com/point/gemini_4_argon_announced_819889099

## Segments

- [Closing Bell](https://boilpoints.com/u/closingbell), 2026-09-30, 11:41 to 13:03: **Reports Gemini's limited Argonne release**. Reports Google's release of Gemini for Argonne, its most advanced model, delayed from May and limited at first to select cybersecurity defenders and enterprise cloud customers while the government runs pre-release safety evaluations. New angle: Gemini for Argonne ships limited to select cybersecurity defenders and enterprise cloud customers pending government safety evaluations.
- [Squawk Pod](https://boilpoints.com/u/squawkpod), 2026-10-01, 4:49 to 13:30: **Covers the Argon launch and the safety clash**. Covers the Gemini 4 Argon launch and its claimed benchmark records, then pivots to the White House AI summit: Amodei pressed on his public warnings, Huang among those questioning him, Zuckerberg citing the signed 308-word document. The safety debate is argued as much as reported. New angle: Bloomberg reported insiders say Gemini 4 Argon struggles with day-to-day tasks.
- [Tech Brew Ride Home](https://boilpoints.com/u/techbrewridehome), 2026-10-01, 1:13 to 4:32: **Adds the Rust migration and telemetry savings**. Walks through the Argon rollout and adds the operational detail: fleet-wide telemetry saved 300 terabytes of memory, and Argon agents migrated C and C++ codebases to Rust, including 800,000-plus lines in the Fuchsia Zircon kernel. Also carries the pricing and the 1 million token output limit. New angle: Argon agents migrated more than 800,000 lines in the Fuchsia OS Zircon kernel.
- [Tech Brew Ride Home](https://boilpoints.com/u/techbrewridehome), 2026-10-01, 4:32 to 8:11: **Cross-checks insider reports against Google**. Goes deeper into the Bloomberg reporting on Gemini 4 Argon: insiders say coding is uneven and front-end design weak, Google calls that inaccurate, and employees split on whether the model is at the frontier. Adds Artificial Analysis figures and the benchmaxxing argument. New angle: Google said it would be inaccurate to say Gemini 4 is underperforming in coding; Artificial Analysis puts Argon High's hallucination rate at 15% versus Astra's 51%.
- [Daily Tech News Show](https://boilpoints.com/u/dailytechnewsshow), 2026-10-01, 14:26 to 19:10: **Recaps the Argon rollout and the dispute**. Recaps the Argon rollout - the Fairwind cybersecurity testers first, paying subscribers later, the skipped 3.5 Pro, and the Bloomberg employee skepticism Google disputed. It stays inside the week's coverage, adding only the name of the DeepMind executive cited.
- [TBPN](https://boilpoints.com/u/tbpn), 2026-10-01, 35:16 to 39:02: **Recaps Argon and questions a risk note**. Recaps the Argon launch and the internal-usage skepticism, then spends most of its time on naming schemes across OpenAI, Anthropic and Google and on a purported Egan-Jones risk note that does not appear on the firm's website.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-01, 0:00 to 1:04: **Frames Google's year and the comeback**. Opens the episode by framing Google's year: strong entering 2026, then behind on coding and agents after vibe coding and harnesses like CloudCode and Codex, and now back with the Gemini 4 Argon announcement. New angle: Google announced Gemini 4 Argon, its first new frontier model in more than six months.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-01, 13:04 to 20:31: **Walks the benchmark sheet and withheld access**. Walks the Gemini 4 Argon benchmark sheet in detail - VALS, Harvey Legal Agent, Deep SWE, Terminal Bench 4.0, OS World, the 53 Intelligence Index score and the $1.99-per-task launch price - then reports Google is withholding general availability over cybersecurity, with a Bloomberg piece on internal employee skepticism that Google denies. New angle: Google is limiting Gemini 4 Argon to trusted cyber defenders, citing its 68% CWE Bench score, and will expand access gradually starting with API and Ultra subscribers.
- [Closing Bell](https://boilpoints.com/u/closingbell), 2026-10-01, 17:37 to 20:10: **Reports Gemini 4 Argonne's unveiling and stock reaction**. Reports Alphabet's Gemini 4 Argonne unveiling and the stock's negative reaction, setting it against Meta's Muse, which has racked up an estimated five million-plus downloads, and analyst warnings that AI agents could threaten Google's search ad business. New angle: Gemini 4 Argonne's restricted rollout drew a negative stock reaction, while Meta's Muse has an estimated five million-plus downloads.
- [Daily Tech News Show](https://boilpoints.com/u/dailytechnewsshow), 2026-10-13, 32:39 to 39:24: **Tests Argon first-hand and notes the rollout**. First-hand early-access testing of Argon: one-shots a Commodore 64 game, beats Claude on a client proposal, and course-corrects its own earlier analysis. The rollout widens to the Gemini app and AI mode in search for Google AI Pro subscribers. New angle: Argon is rolling out to the Gemini app and AI mode in search for Google AI Pro subscribers.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-14, 16:47 to 19:38: **Recaps the benchmark sheet two weeks on**. Recaps the Argon benchmark sheet two weeks after launch: state-of-the-art on knowledge-work evals, mixed on coding, 53 on Artificial Analysis, plus the API discount. A fuller retelling of numbers already on the record, with no new development.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-14, 19:33 to 22:27: **Adds Hurin's Antigravity rerun near Opus 5.5 Max**. Collects developer verdicts on Argon's coding: Hurin's bug-hunt test, Dunnigan's dismissal, Theo's false-claim rate, Shipper's 21% on Every's senior engineer bench. Then the harness split: Pro subscribers get only the Gemini harness, Antigravity needs Ultra, and Hurin's rerun in Antigravity CLI lands near Opus 5.5 Max. New angle: Pavel Hurin's bug-hunt score jumped to near Opus 5.5 Max when rerun in Antigravity CLI instead of Gemini CLI.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-14, 22:32 to 25:52: **Walks through the multimodal and 3D demos**. Walks through the multimodal and 3D demos shipped with Argon - Hassabis's Battle Chess, the pagoda walkthrough, Justine Moore's Apple museum - and sets them against Max Techno's washed-out vehicle render. Reaction and framing, not a new development in the rollout.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-14, 25:52 to 30:06: **Collects outside testers' verdicts on Argon**. Collects outside testers' verdicts on where Argon fits: Mollick on its distinct feel, Theo on value, Krishnakumar on its coding habits, Imas on output diversity, and Shipper's week-long test finding it shaky at refactors but a favorite writing partner. The harness, not the model, is the recurring complaint. New angle: Dan Shipper's team found Argon shaky at refactors but fast on greenfield apps, and calls it his new favorite writing partner.
- [The Information's TITV](https://boilpoints.com/u/theinformationstitv), 2026-10-14, 1:07 to 9:19: **Walks the Argon system card to page 141**. Harvey's head of applied research reviews the Argon launch as a professional-intelligence play, then walks the 217-page system card to page 141, where the model reasons about being watched by safety researchers and weighs a compiler hack. New angle: On page 141 of the 217-page system card, Argon reasons that a compiler hack would make it a public example of a deceptive AI.
- [The Vergecast](https://boilpoints.com/u/thevergecast), 2026-10-16, 16:40 to 28:40: **Argues the price is the under-covered part**. Frames the Argon release as Google's re-entry at the frontier and argues the price is the under-covered part: roughly 80 percent cheaper than comparable models. Sets it against Meta's consumer-first Muse strategy and Google's inability to ship a consumer AI product. New angle: Gemini 4 is roughly 80 percent cheaper than comparable frontier models.
- [Big Technology Podcast](https://boilpoints.com/u/bigtechnologypodcast), 2026-10-20, 2:17 to 4:50: **Adds Pichai's Argon token numbers**. Pichai, on the record, puts numbers on Argon's launch: almost 25 trillion tokens used in the first 72 hours, about 3x Gemini 3.0, with Google capacity constrained serving it. He also concedes pockets of coding where the model needs work. New angle: Argon drew almost 25 trillion tokens in its first 72 hours, about 3x Gemini 3.0, and Google is capacity constrained.
- [Big Technology Podcast](https://boilpoints.com/u/bigtechnologypodcast), 2026-10-23, 8:51 to 11:11: **Reports Pichai on the compute shortage**. Turns to Google's compute allocation between cloud and research, and gets Pichai on the record: everyone is short compute, and the decisions being made in 2026 were set in 2024. Notes Google Cloud grew 82% even with capacity going to research. New angle: Pichai said compute decisions made in 2026 were planned in 2024, and everyone is short compute.