Google stands by Gemini 4 performance as some claim it ‘struggles’ in real-world use

Google recently unveiled its latest flagship model, Gemini 4, boasting impressive benchmark scores. However, Bloomberg reports that some Google employees have discovered that the model's real-world performance may not always match these scores. The report cites internal Googlers who found that Gemini 4 "does less well when employees actually put it to work" and struggles with certain coding tasks.

Despite these claims, Google maintains that the model is not underperforming, particularly in areas like coding. The company points to comments from Koray Kavukcuoglu, head of Google DeepMind, who stated in late September that "we are always gonna be at the frontier." While some employees believe Gemini 4 is at the cutting edge, others are concerned it may still lag behind competitors like Anthropic and OpenAI.

Gemini 4 Argon, the new model, brings upgrades such as an expanded output limit of up to 1 million tokens, and Google has also highlighted its "highly capable performance in cybersecurity defense." It will be available at a cost of $4 per 1 million input tokens and $20 per 1 million output tokens, with introductory pricing at half of that rate.