Some Google employees have found that Gemini 4 performs less well in day-to-day use than its benchmark results suggest, particularly on certain coding tasks, according to a Bloomberg report that 9to5Google summarized Thursday. Google disputed the account, saying it would be wrong to describe the model as underperforming in areas such as coding.

What Happened

According to 9to5Google's summary of the Bloomberg report, the concern among some staff is that Gemini 4 does worse once employees actually put it to work than its published scores imply. The report singled out coding as an area where the model struggles with some tasks. Internal opinion was not uniform: 9to5Google said some employees agreed the model had reached frontier-level performance, while others worried it would still trail models from Anthropic and OpenAI.

Google said it would be "inaccurate to say" that Gemini 4 is underperforming in coding and similar areas, and it pointed to earlier remarks by Koray Kavukcuoglu, who leads Google DeepMind, that the company will always be at the frontier of AI development.

RelatedAnthropic Prices Claude Haiku 5.5 at a Tenth of Its Predecessor for Most API Requests

The report landed a day after Google introduced Gemini 4 Argon, the first model in the family. When it unveiled Argon, Google said the model outperformed OpenAI's GPT-6 Astra in many key areas and highlighted its strength in cyber defense. 9to5Google noted that Argon's output limit rises to 1 million tokens and listed its standard API price at $4 per million input tokens and $20 per million output tokens, with introductory pricing at half those rates.

Why It Matters

Coding is one of the most commercially important uses of frontier models and one of the most competitive. Developers and enterprise buyers compare models on how well they handle real repositories, long-running tasks and messy specifications, not only on leaderboard numbers. A report that a model's own builders see a gap between test results and practical work goes to the heart of how Google is marketing Argon, which it pitched for long-running software engineering and vulnerability hunting.

The criticism also fits a pattern that has followed many recent frontier launches: benchmark suites, often published by the developer itself, can overstate how a model behaves on unfamiliar or open-ended problems. Independent comparisons of Argon were already mixed before the Bloomberg report. As The Primary Wire reported Wednesday in "Google Opens Gemini 4 Argon to Cyber Defenders Before a Wider Release," VentureBeat counted 18 benchmarks Google disclosed and found Argon first on 12, but behind GPT-6 Astra by about 10 points on the FrontierSWE v2 agentic coding test.

Limited Access So Far

Argon's rollout makes the dispute hard to settle quickly. Google released the model first to vetted cyber defenders in its Fairwind security program and said broader access would start with paid API customers and Google AI Ultra subscribers, without giving a date. That means most of the developers who would test the coding claims in practice cannot yet use the model.

Bloomberg's account relies on employees' impressions rather than published test data, and 9to5Google did not report any specific internal evaluations or failure rates. Google, for its part, did not release new figures with its rebuttal.

What to Watch

The clearest test will come when Argon reaches paid API users and third-party evaluators can run it on their own codebases. Developers will be watching how it compares with GPT-6 Astra and Anthropic's latest models on agentic coding work, and whether Google publishes additional evaluations to answer the internal criticism. Any change to the release timeline or the introductory pricing would also signal how confident Google is in the model's position.