Gemini 4 Argon Flops in Coding Tests: 51% Hallucination Rate, Trails Claude Opus 5.5

The author, a longtime Gemini 3.8 Flash user, gains access to Gemini 4 (internal codename Argon) via the Antigravity CLI, where the model appears as Gemini 4 Argon with adjustable Thinking Effort up to High. A notable quota improvement: Argon and Gemini 3.8 Flash draw from separate pools, unlike Claude Opus 5.5 which consumes 5 hours of quota per query.

Complex Animation and Video Generation Test

Using the same prompt that recently showcased Claude Opus 5.5's code animation prowess (a hand-drawn boiling-line matchbox ignition), the author tasks Argon with the identical challenge. With High thinking enabled, Argon takes over 30 minutes to produce a result described as rough and barely passable, consuming over 30% of the weekly Argon quota in a single run, leaving almost no room for iteration.

Blind Architecture Audit on Production Codebase (Theo's Evaluation)

Frontend influencer Theo conducts a rigorous blind audit on his production-grade T3 codebase. The results expose Argon's severe shortcomings:

Claude Opus 5.5 leads with 8.40/10, only 2 false claims out of 42 (5% hallucination rate).

Claude Sonnet 5.5 scores 7.88 with 9% hallucination.

GPT-6.1 Sol and GPT-6 Astra achieve 7.66 and 7.11 respectively.

Gemini 4 Argon collapses to 3.95, bottom of the flagship tier, with 47 fabricated assertions out of 92 system claims — a 51% hallucination rate.

Argon fails to grasp real project architecture, invents non-existent interfaces, and confidently proposes refactors that would break the build. Theo's verdict: "I was genuinely hopeful for Gemini 4 Argon. There are some good parts, but generally speaking, this model should be avoided."

Three.js 3D Rendering Comparison (Bhavani's Test)

Developer Bhavani prompts four models to build the Eiffel Tower purely in Three.js code. Contenders: Gemini 4 Argon, Claude Opus 5.5, GPT-6 Astra, and Kimi K3.

Argon's tower exhibits distorted base structure, incorrect proportions, and lacks mechanical and spatial symmetry.

GPT-6 Astra and Kimi K3 produce detailed, layered towers closely matching the real landmark.

Claude Opus 5.5 maintains its consistent solid performance.

Argon again shows a disconnect between spatial reasoning and code precision in fine-grained geometry tasks.

Google's Internal Benchmarks vs. External Reality

Google's official blog touts impressive internal metrics for Argon:

Over 100,000 Alphabet engineers rely on Argon monthly.

Internal agent code output surged 35% in four weeks; post-commit rollback rate dropped 30%.

YouTube frontend comment-section optimization halved CPU overhead and cut end-to-end latency by 20 ms.

Wing drone simulator received 18 high-difficulty aerodynamic modeling improvements.

Decades-old Fortran scientific code fully rewritten to modern Rust, yielding 1.8× performance gain and memory safety.

Security audit uncovered 4 hidden permission vulnerabilities previously missed by two other top models.

72.4% of cross-system legacy bugs cleared automatically within 3 hours.

The author highlights the critical context: these results stem from Google's world-class standardized engineering stack — monorepo, comprehensive runtime telemetry, thousands of pre-commit tests, and deeply customized Custom Skills tailored to internal workflows. External developers face heterogeneous tech stacks, tangled third-party dependencies, and vast untyped legacy codebases. Without Google's guardrails, Argon's context-awareness flaws surface dramatically: it over-hallucinates, fabricates APIs, and delivers plausible-looking but broken solutions.

Fairly, Argon retains Google-grade strengths in massive-context information extraction, long technical document synthesis, and multimodal analysis. However, for day-to-day production coding in uncontrolled environments, the evidence strongly favors Claude Opus 5.5 and, surprisingly, GPT-6 Astra and Kimi K3.