Google's Gemini 4 Argon unleashes 1M output tokens

For the past year, AI labs have raced to ingest massive libraries of text, but the moment you asked a model to write back more than a few thousand words, it hit a wall and lost the plot.

Google answered back with Gemini 4 Argon, its newest frontier model built specifically for heavy lifting and extended problem-solving. Its headline breakthrough is an industry-first 1 million output token limit, shattering the previous 64K ceiling and letting the model produce roughly 700,000 words in a single, uninterrupted run.

Argon backed up the size with raw horsepower, putting up a chart-topping 77.9% on the DeepSWE v1.1 coding benchmark, scoring 51.3% on Zapier's AutomationBench, and outperforming GPT-6 Astra and Claude Opus 5.5 on 13 of 19 internal head-to-head tests.

Rewriting Google from the inside out

Instead of rushing to a public chat interface, Google first turned Argon loose on its own infrastructure.

Inside the company, Argon agents have been tasked with migrating sprawling C/C++ codebases to memory-safe Rust, including components spanning up to 800,000 lines in the Fuchsia Zircon kernel. In libgav1, an open-source video decoder, Argon ran automated profiling experiments, rewrote 32,000 lines of SIMD instructions into clean, safe Rust, and delivered 2.7x faster performance.

Across server farms, teams of Argon agents combed through fleet telemetry to pinpoint memory leaks, freeing up over 300 TiB of data center memory upon rollout, with long-term savings projected to top 1 PiB.

Argon even helped quantum researchers optimize circuit bottlenecks, beating published research baselines by 40% in just minutes.

Unshackled guards for cyber defenders

Google is rolling out Argon cautiously. The model is currently restricted to vetted cybersecurity experts through Google’s Fairwind Program, with access to paid API tiers and Google AI Ultra subscribers slated for later.

Because Argon was explicitly trained to autonomously discover, verify, and patch zero-day vulnerabilities, Google is shipping an unguardrailed version specifically to trusted defenders so they can exploit its full forensic depth.

Cloud security firm Wiz has already put Argon to work in its Scan for Good project, uncovering a severe data leak vulnerability in hospital software worldwide that prior frontier models failed to notice.

Early independent benchmarks like Arena and Artificial Analysis place Argon right in the top tier alongside GPT-6 Astra. However, not everyone inside Mountain View is sold: Bloomberg reported internal whispers that Argon looks stronger on synthetic tests than in day-to-day coding, an accusation Google firmly denies.

Pricing begins with an introductory rate of $2 per million input tokens and $10 per million output tokens, before stepping up to $4 and $20.

Why it matters

After a quiet 2026 marked by the cancellation of Gemini 3.5 Pro and months of relying on mid-tier Flash models, Google is definitively back in the frontier heavyweight class.

Pushing output capacity to 1 million tokens fundamentally shifts what autonomous agents can attempt. Until now, coding assistants were confined to small functions because context decay and output throttles severed long-horizon plans. A million output tokens provides the reasoning runway needed to rewrite entire architectural stacks or execute complex multi-day audits in one sustained trajectory.

For enterprise developers and security leaders, this moves AI from autocomplete to genuine delegation. But the trade-offs are real: a million-token generation will be costly and slow, and debugging an agent that wanders off track deep into an hour-long run will test anyone's patience.

Google’s defensive security angle is a smart wedge. Defensive cyber teams have the budget and the tolerance for complex verification, making them the ideal proving ground before the rest of us get to find out whether Argon's real-world stamina matches its benchmark hype.