Gemini 4 Argon Writing Test: 173K Book, 1M Context, Beats GPT-6/Claude

The author, an AI developer known as 鲁工, conducts an intensive hands-on evaluation of Gemini 4 Argon's long-form writing capabilities using the Antigravity CLI. Over a week of high-intensity testing, they conclude that Gemini 4 Argon currently occupies a unique tier for extended writing and textual control.

Book-Writing Experiment: 《姜维传》

The centerpiece test was a complete biography of the Shu Han general Jiang Wei (姜维). The prompt required the model to first compile a 20,000-character source anthology from four classic Chinese historical texts — 《三国志》 with Pei Songzhi's annotations, 《华阳国志》, 《晋书》, and 《资治通鉴》 — then produce a 12-chapter, 173,000-character manuscript. The entire workflow ran inside Antigravity CLI, which manages the 1-million-token unified context window (input, internal thinking tokens, and output share a single budget) and the continuation protocol for long outputs.

After the initial draft, the author used DeepSeek V4.1 Flash as a reviewer for three rounds of critique and revision, then exported a print-ready PDF. Screenshots in the article show the CLI command, the generated chapter outline, the PDF layout, and several polished passages the author describes as "delicate and nuanced" beyond their own writing ability.

Why Gemini 4 Argon Writes More Naturally Than GPT-6 Astra or Claude Opus 5.5

The author identifies two structural advantages rooted in Google's training pipeline:

Pre-training corpus: Decades of scanned Google Books publications give the model native exposure to professional book-length narrative rhythm and vocabulary texture, avoiding the "official summary flavor" common in models trained mainly on web pages and synthetic data.

Post-training alignment: The 1M-token budget is shared across input, thinking, and output. Official blog posts confirm that the reinforcement learning phase does not suppress the internal chain-of-thought. This grants the model ample reasoning space to structure long works and internalize negative style constraints (e.g., "no parallelism, no dash-heavy insertions, no formulaic section headers") without defaulting to the high-scoring RLHF templates (general-specific-general, bullet lists, rhetorical flourishes) that plague Claude Opus 5.5 and GPT-6 Astra even when detailed style guides are provided.

Benchmark Evidence: Surge AI Hemingway-Bench

The article cites Surge AI's Hemingway-bench evaluation, which involved 5,000+ double-blind comparisons judged by professional screenwriters and senior editors. Key findings:

Automated EQ-Bench scores correlate poorly with human literary judgment — only 43% agreement — because they reward dense metaphor and ornamentation.

Under human blind evaluation and LMArena text arena, Gemini models rank highest for restrained, detail-rich prose under strict constraints.

Long-form quality degradation rate: Gemini 4 Argon 0.029 vs. GPT-6 Astra 0.050 (nearly half).

Critical Failure Mode: Confident Hallucination of Citations

The author warns that Argon's high confidence becomes a liability when it exceeds the provided source material. During the first review pass, DeepSeek caught a fabricated citation: a passage attributed to 《三国志·郤正传》 Pei Zhu was actually from 《晋书》. The model invented a plausible-looking academic reference rather than admitting uncertainty. The author stresses that AI-generated text for serious contexts — papers, patents, books, public releases — must be human-reviewed: "Writing feels great for a moment, but mistakes lead to disaster."

Supporting Data Points from Google

Google AI Studio lead Logan Kilpatrick and CEO Sundar Pichai released 72-hour post-launch metrics: ~25 trillion tokens processed, ~96 million tokens/second average throughput (3× Gemini 3.0 same-period usage). Logan's launch thread listed writing capability as the top excitement point, followed by raw knowledge, coding endurance, and the output ceiling jump from 64K to 1M tokens.