Google's global rollout of Gemini 4 Argon, a frontier model that expands its output token limit to 1 million tokens and tops specialty benchmarks in coding, robotics, and long-context reasoning while posting only a tied score on broad independent intelligence rankings.

Google's headline technical change in Gemini 4 Argon is not a new architecture but a sixteen-fold jump in how much the model can write in a single turn: the output token limit rises from 64,000 to an industry-leading 1 million tokens ^[1]. For long-horizon workflows - multi-file refactors, hour-long video analysis, extended agent plans - the output ceiling is often the real bottleneck, not the model's reasoning quality: a model that reasons well but gets cut off mid-task still fails the task. Argon's benchmark results point to exactly this mechanism at work. It sets a new state of the art on DeepSWE v1.1 at 77.9%, a long-horizon software engineering benchmark, and on LVBench at 91.7%, a long video understanding benchmark ^[1]. Both wins land squarely in categories where a 64K-token ceiling would previously have forced the model to stop before finishing.

Where Argon competes on narrow, task-specific benchmarks, it wins outright. It ranks first on Vals AI's GDP-weighted Vals Index, scoring 68.90% as of October 12 ^[2]; it ties for first place on CWE-bench v1, the benchmark for remediating security vulnerabilities, with a top score of 68% ^[1]; and in a robotics evaluation using pure reasoning, it scores 87%, outperforming every other frontier model ^[3]. But on the broader, generalist measure that cross-lab buyers tend to cite, Artificial Analysis's Intelligence Index, Argon only ties GPT-6 Astra and sits behind Claude Opus 5.5 ^[2]. The one place Argon clearly separates itself in that broader evaluation is factual reliability: Artificial Analysis puts its hallucination rate at 15% on AA-Omniscience, versus 51% for GPT-6 Astra ^[2]. That split - specialty-benchmark leader, generalist tie, standout calibration - is the real shape of Argon's competitive position: dominant where the task is narrow and well-defined, merely even where it is not.

Argon launches at $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input tokens ^[2]. That pricing is explicitly temporary: it expires November 30, 2026, after which the standard public API price doubles to $4 per million input tokens and $20 per million output tokens ^[2]. The bigger real-world cost risk is on the output side rather than the sticker price. One developer benchmark found Argon generates roughly 62,000 output tokens to reach the same score that GPT-6 Astra reaches with about 27,000 tokens ^[4]. At output pricing, a model that is more than twice as verbose per task can erase most of the headline discount - the actual bill for long-horizon workloads depends far more on how many tokens a task consumes than on the quoted per-token rate.

Google first unveiled Gemini 4 Argon on September 30, 2026, rolling it out only to a set of trusted cyber defenders through the Fairwind Program ^[1]. Roughly two weeks later, on October 13, 2026, Argon became globally available to developers and enterprises via the API in AI Studio and the Gemini platform ^[5]. The launch is widely read as Google's attempt to re-establish AI leadership after a DeepMind leadership shakeup and release delays ^[6], with Alphabet's stock posting a mixed reaction to the news ^[7]. JPMorgan framed the stakes directly, writing that Argon needs to help Google "re-establish itself at the frontier and as an AI leader for both the industry and investors" ^[7]. That framing only makes sense against the pace of the last year: Gemini 3 Pro, the prior flagship, launched in November 2025 ^[8], putting Argon's arrival roughly eleven months later into a market that had not been standing still.