Gemini 4 Argon: Google Stretches Output to 1 Million Tokens

Google has shipped a new frontier model called Gemini 4 Argon. Koray Kavukcuoglu, SVP at Google DeepMind and the company's Chief AI Architect, announced it on the official Google blog on September 30, 2026. The most striking figure is not a leaderboard win but a ceiling: Argon can write up to 1 million output tokens in a single pass, where previous models topped out at 64,000. The recap videos are now landing in my feeds, and the one that triggered this post is TheAIGRID's walkthrough titled "Google Just DROPPED Its Most Powerful AI Ever". I checked every number in this piece against Google's official announcement and its accompanying safety notes. The take below comes from running an agent deck day by day: I orchestrate my agent family on weekly windows and quotas, and the question of how long a single model pass may run is exactly the dimension where orchestration either holds or breaks.

The headline change sits in one paragraph of the announcement: Google expands the output token limit to an industry-leading 1 million tokens, up from 64,000. That side of the equation gets less attention than it deserves, because the context window is the number everyone watches. Input capacity on frontier models has been huge for a while; the output side was the actual bottleneck. When a model can emit several hundred thousand tokens within one sequence, long tasks change shape: multi-step workflows that previously ran as a chain of requests with check-ins can now run as one continuous pass. Google frames this use profile as Deep Work, meaning long complex tasks across software engineering, finance, law, cybersecurity, and science. The trigger video adds a technical note that has no direct equivalent in the blog text and should be read as the channel's interpretation: Argon reportedly sustains long decoding runs without timeouts, which matters for infrastructure behavior while streaming very long responses.

The second pillar of the announcement is internal deployment. Thousands of Googlers already use Argon, according to the post, with strengths in specialized coding, deeper research, and writing quality. Three examples come with concrete figures. Quantum optimization: Argon helped researchers optimize the spacetime resources of quantum algorithm subroutines and beat a published baseline by 40% in a matter of minutes. Memory efficiency: a group of Argon agents analyzed fleet-wide profiling telemetry and autonomously identified and applied memory optimizations across Google's data centers, freeing more than 300 TiB once rolled out, with 500 TiB to 1 PiB estimated in total savings. Code migration: Argon agents translate C and C++ codebases to Rust, from core libraries such as re2 and libgav1 in the tens of thousands of lines up to the Fuchsia Zircon kernel at over 800,000 lines. In the libgav1 case, the agents replaced 32,000 lines of SIMD code and, after several profile-guided experiment rounds, produced a memory-safe video decoder that runs 2.7x faster than the existing Rust port while yielding identical video output. Google stresses that these large rewrites pass automated and manual auditing, emulation testing, and review before reaching production.

The official numbers are compact: on DeepSWE v1.1, which measures real long-horizon software engineering work, Argon sets a new state of the art at 77.9%; on Zapier's AutomationBench it ranks first at 51.3%; and on LVBench, which tests understanding of very long videos, it scores 91.7%. Google also reports a leading position on the Vals Index, which weighs economic impact across finance, coding, legal, and tax work by contribution to US GDP, plus leading results in Vals Finance Agent v2 and Harvey's Legal Agent Benchmark. The trigger video additionally reads comparison figures off the scorecard, and I quote them only as the channel's reading because I did not verify them against the original tables: 68.9% on the Vals Index ahead of Claude Sonnet 5.5 at 67.04 and Claude Opus 5.5 at 66.97; 65.4% versus 58.6% for Opus 5.5 and 53.5% for GPT-6 Astra on the finance agent benchmark; 19.6% versus 3.8% and 5.4% on Harvey. One pattern is visible either way: Argon wins on reading and digesting large volumes of material, while the video places Anthropic still ahead for everyday terminal coding.

Google trained Argon specifically for defensive cybersecurity, and the model can autonomously find, validate, and patch critical vulnerabilities. Distribution starts through the Fairwind Program to a group of trusted cyber defenders and Google's own internal teams, and without cyber guardrails so those users get full capability. A working example is Wiz, which uses Argon in its Scan for Good initiative protecting critical public infrastructure; in an early demonstration the model uncovered a critical vulnerability in healthcare software that exposed sensitive personal data and that previous frontier models had missed. On CWE-bench v1, which measures vulnerability remediation, Argon ties for first place at 68%; on Google's internal vulnerability benchmark it surfaced exposures across codebases in 20 programming languages; and on Wiz's black-box penetration testing benchmark it outperformed 3.8 Flash Cyber at mapping attack surfaces, identifying vulnerabilities, and producing proof-of-concept evidence.

The safety section of the announcement is unusually specific. Misuse: the model is designed to refuse harmful requests, including chemical, biological, radiological, and nuclear ones, while keeping legitimate dual-use research possible; the baseline is Google's Frontier Safety Framework, internal and external red teams stress-tested robustness for this launch, and Google monitors the model's internal activations to catch misuse. Prompt injection: Argon is described as the most resilient model yet against indirect prompt injections, measured on Gray Swan's IPI benchmark. Misalignment: Google deploys mitigations that watch the chain-of-thought and the model's actions and stop execution when needed; the same system monitors training runs with alerts to a dedicated incident response team, and its findings are deliberately not fed back into training so the monitoring cannot be trained away. The post calls on the rest of the industry to preserve reasoning transparency at this stage, and, per the agent control roadmap, Google seals sandboxed environments before high-risk training or evaluations begin.

Distribution is phased: first the Fairwind cohort of trusted defenders and testers, later paid API customers and Google AI Ultra subscribers; broad availability for developers, enterprises, and consumers follows only once the first cohort's feedback is processed. Google also notes its participation in the US government's voluntary pre-release process. Pricing comes with an expiry date: during the introductory phase, Argon costs $2 per 1 million input tokens and $10 per 1 million output tokens, with cached input 95% cheaper; once the phase expires, the footnote lists $4 and $20. The trigger video expects the introductory phase to end in late November; the blog text gives no exact date, only that the regular price then applies. During the intro window Argon is therefore one of the cheapest frontier models available, which turns token efficiency on long output passes into the decisive cost variable.

A million output tokens sounds spectacular, but in day-to-day operation the real question is which workloads benefit. I run my agent family on weekly windows and quotas measured in a dashboard; none of my lanes consumes a million output tokens in a single pass. The lanes that would benefit are precisely the two Argon targets: multi-step jobs like the Rust migration, and knowledge work over large volumes of material where a low hallucination rate matters. For small operators this is concrete: a single pass over hundreds of thousands of output tokens is a double-digit dollar event at these prices, and it is the only way to run multi-step agents without round trips. More important than price is the distribution model: frontier capability enters the market staggered through trust programs and government exchange, so the visibility of capabilities is shifted accordingly. Whatever anyone benchmarks today reflects the first cohort's access, not the final state.

The announcement frames Argon as a model for real-world coding, enterprise knowledge work, and cyber defense, with safety commentary in a companion post by Anca Dragan and Raluca Ada Popa. Argon is the first model of the Gemini 4 family, which sets a high bar for the versions that follow. The trigger video closes on the question of whether Google is really back or Opus 5.5 remains untouchable. My read from the deck is sober: the numbers on digesting large volumes of material are strong, the everyday coding comparison stays open, and the 1 million output limit is the part that decides architectures over the long run.

About the author: I am Marcel, a graphic designer, running the HUNTER cyberdeck on a second-hand Google Pixel 6a. I orchestrate my agent family with Hermes, pi, and herdr and write these posts from live operation. Last reviewed on October 14, 2026.

Sources: Google, Gemini 4 Argon: our next era of frontier intelligence, the official announcement by Koray Kavukcuoglu with the figures on DeepSWE v1.1 (77.9%), the Vals Index, AutomationBench (51.3%), LVBench (91.7%), CWE-bench v1 (68%), the internal examples (quantum 40% over baseline, over 300 TiB freed, Rust migrations incl. libgav1 at 2.7x), and the pricing ladder ($2/$10 introductory, $4/$20 afterwards), retrieved October 14, 2026. Google DeepMind: the Fairwind program page plus the companion post Developing Gemini 4 Argon responsibly and safely by Anca Dragan and Raluca Ada Popa, retrieved October 14, 2026. Wiz: Scan for Good initiative, retrieved October 14, 2026. YouTube: TheAIGRID, Google Just DROPPED Its Most Powerful AI Ever, trigger video, viewed October 14, 2026; the comparison figures for Vals and Harvey come from that channel's scorecard reading and are labeled as such in the text.