Gemini 4 Argan Is Gorgeous on the Surface and Unreliable Underneath

Google's first big Gemini release in eight months tops benchmark charts and builds slick landing pages. Put it inside a real codebase, though, and it gets more wrong than right.

The wait was long. Since the 3.1 Pro release back in February, Google had shipped nothing but Flash models while its rivals rolled out new flagships every month or two. So when Gemini 4 Argan finally landed this week, after a teaser campaign of more than a week, developers were ready to be impressed.

One of the people with early access was the developer behind T3 Code, who put the model through his own battery of tests and published a lengthy breakdown. His conclusion was blunt. Argan has real strengths, but as a coding model it is, in his words, one of the worst releases of the year.

Here is what the testing actually showed.

The price tag is misleading

On paper, Argan looks like a bargain at $2 per million input tokens and $10 per million output. Next to Opus 5.5's $5/$25 rate, that seems like an easy win.

It isn't, for three reasons. First, the model is token hungry. On Artificial Analysis's intelligence index, Argan scores above GPT-6 Astra and GPT-6.1 Soul but clearly below Opus 5.5 on high, and it still costs more to run than Opus on that same index.

Second, prompt caching barely works. The reviewer saw cache hit rates of roughly 70 to 80 percent, compared with 99 percent or better from other models. That means the big discount on cached input rarely kicks in. He allowed that this could be a bug in Google's tooling, but argued it's Google's job to fix, not his.

Third, the current price is temporary. Google is calling it introductory, and it doubles to $4/$20 on December 1. Raising prices after launch has become a Google habit, and it makes it hard to commit to building on the model.

Where it genuinely shines

Front end design. Gemini models have long been good at landing pages, and Argan keeps that up. Paired with Anthropic's official design skill, its homepages looked polished enough to pass as human work, with functional interactive sections. Without the skill, it falls back to cluttered layouts and the same tired top navigation. The reviewer still ranks Fable 5.1 as the best front end model he has used, but Argan is a solid option for a first draft.

A perfect score on a weird benchmark. On Skatebench, his private test of whether models can correctly name skateboard tricks, Argan scored 100 percent. No model had done that before. Opus 5.5 on max reaches 79 percent. It was also fast and economical with tokens, which is unusual for Gemini.

Google's own data. The Gemini API has privileged access to YouTube, which makes it one of the best ways to pull transcripts and video data. For anyone working with YouTube content, that alone keeps Gemini in the pipeline. Gemini has also historically led on vision tasks like labeling parts in a photo, so it's worth benchmarking for those.

The 3D test exposed a deeper problem

His standard 3D benchmark is a small aquarium game he has every new model build. Argan's first attempt had a broken water surface, a mediocre submarine model, and a bug where food sank to the bottom, never expired, and left the fish staring at it until they starved.

A second pass fixed most of that. But it also revealed something telling. The model added an attack power-up that expires before the next wave of monsters arrives, so it can never be used. Nothing about it was a hallucination. It just wasn't thought through. The reviewer described this as missing "big model smell," the sense of understanding a whole system rather than writing code that merely looks reasonable.

Benchmarks say Argan rarely hallucinates. Real codebases say otherwise.

Hallucination: the numbers don't match the marketing

Artificial Analysis credits Argan with a 15 percent hallucination rate, the lowest among high-scoring models. But that benchmark rewards declining to answer, and Argan's overall score there drops to fourth because it simply gets fewer things right.

The reviewer's own tests told a very different story. He had several models audit real codebases, then had every claim checked and graded by a panel of judges.

Scores out of 10, from the reviewer's own judged audits.

In the Lakebed audit, roughly 70 percent of Argan's findings were wrong. In the architecture review, it produced more false claims than Muse Wonderful Full produced true ones. The reviewer's advice: if you use Argan for reviews at all, run its output past a more reliable model before anyone acts on it.

A 1,200-line fix for a 30-line problem

The final test was practical. T3 Code users on Windows were hitting SQLite connection instability, and the reviewer already knew the fix: separate read pools with a single writer connection. He wanted to see whether Argan could get there.

It did identify the read/write split. But asked to open a PR for its top priority, it took 30 minutes and produced roughly 1,200 lines of code for what should have been a small change. The PR ignored the repo's formatting rules, failed both lint and CI checks, and never addressed comments from the review bots.

When Opus 5.5 was asked to review the work, it found that Argan had fixed the wrong half of its own diagnosis. Its new reader pool still ran on the main thread, so slow reads would block exactly as before. It also missed two existing PRs that already solved the problem. Opus then wrote the correct fix in 21 lines, which has since been merged.

The ecosystem is part of the problem

Beyond the model itself, the reviewer was frustrated with Google's tooling. Its agent integration is unusually awkward, with effort levels baked into model names instead of exposed as a setting, and a large share of T3 Code's bugs trace back to the Antigravity integration. During testing, the model dropped its connection an hour into a run and couldn't resume, forcing a restart with no history.

There is also account risk. A ban from Anthropic or OpenAI costs you one service. A Google account holds email, billing, saved passwords, and potentially a YouTube channel. Google says it now bans only from Gemini products, but the reviewer has seen reports suggesting otherwise.

Google insiders reportedly describe Argan modestly, as something that makes them reach for the older Opus 5 a bit less. There are rumors of a stronger snapshot coming soon. It will need to be a big step up. For now, Gemini 4 Argan is a model that looks the part without reliably doing the work.