Some Google employees say Gemini 4’s strong benchmark results do not consistently carry over to their coding work, according to Bloomberg’s September 30 reporting. The report, citing people with direct access to the effort, adds an internal disagreement to the mixed outside results accompanying Argon’s launch.

Google disputes the characterization. It told Bloomberg that describing Gemini 4 as underperforming in coding would be inaccurate. Bloomberg also spoke to an employee familiar with model development who said internal testing supported its standing among the leading models and rejected claims that it struggles with messy coding tasks. Other employees believe rivals are advancing faster.

Bloomberg separately reports that Google abandoned Gemini 3.5 Pro, which it had planned to release in June. That is a reported cancellation of an earlier model, not a halt to Argon’s rollout.

The public evidence supports a narrower conclusion than either a clean victory or a failed model. Arena’s WebDev ranking places Argon eighth, behind leading Claude and GPT models. Google, meanwhile, describes specialized engineering successes, including a Rust video decoder it says Argon made 2.7 times faster than the previous Rust version.

Those are different kinds of coding work. The employee accounts raise a practical question about consistency, but do not establish a measured failure rate. Access still starts with selected cyber defenders; Google has not given a date for broader availability.

Gemini 4 Argon ranks eighth overall but first at responding to user corrections in Arena’s new results from real agent sessions. At high reasoning, it places ahead of Gemini 3.8 Flash in nineteenth, with a median task cost of $0.62 versus Flash’s $0.58.

Unlike Arena’s head-to-head chat votes, Agent Arena randomly assigns models to handle users’ tasks, then measures feedback and tool behavior. Its steerability signal checks whether a user accepts the agent’s fix after correcting it. Argon also ranks second on explicit user-confirmed success.

The overall score combines those signals with praise versus complaints, recovery from failed commands and attempts to call nonexistent tools. Argon’s +7.92% net improvement is an estimated gain over Arena’s baseline, not the percentage of jobs it completes. Claude Fable 5.1 at maximum effort leads the overall ranking.

The result is preliminary: Argon has 3,417 sessions, and its overall score carries a 95% confidence interval of ±3.24 percentage points. Its $0.62 median uses Google’s introductory $2/$10 input/output rates per million tokens, so it is not a permanent-price estimate.

The useful early signal is about collaboration: Argon looks particularly good at getting back on track when a person intervenes. That is different from demonstrating reliable, unattended work.

Gemini 4 Argon shows a sharp improvement over Gemini 3.8 Flash on Artificial Analysis’s hallucination measure: 15% versus 55%, with both models set to high reasoning. But the accompanying accuracy results show what kind of improvement this is: Argon gets fewer questions fully right, too.

The test, AA-Omniscience, asks 6,000 factual questions across areas including law, health and software engineering, without browsing or other tools. It rewards correct answers and penalizes wrong guesses, while leaving abstentions unpenalized.

The denominator matters. That 15% counts outright incorrect answers as a share of responses that were incorrect, partial or not attempted. It does not mean Argon is 85% accurate. On the separate accuracy measure, Argon answers 50% of all questions correctly, compared with Flash’s 55%.

Together, the results point to a more cautious model: substantially fewer outright wrong answers, alongside more partial answers or abstentions, rather than a jump in factual recall. That can be valuable when a missing answer is better than a false one. It is not evidence that hallucinations have been eliminated in everyday use.