Gemini 4 Scores Well on Benchmarks, but Employees Report Coding Gaps

Bloomberg reported that Gemini 4 performed well on commonly used AI benchmarks but did less well when employees tried it on real tasks, including some coding work. The report captured internal doubts ahead of the model’s release—not a definitive assessment of every capability. Google disputed the criticism, and a same-day report said the company had unveiled Gemini 4 Argon, though it did not give a timetable for general availability. 7510

What Bloomberg reported about Gemini 4

People with direct access to the effort told Bloomberg that Gemini 4 scored well on industry benchmarks but struggled with certain coding tasks in practice. The accounts were anonymous, so the reported gap is best understood as employee feedback rather than a published, independent evaluation of the model. 7

That distinction matters: benchmark results and reports from people using a model on particular tasks are different kinds of evidence. The Bloomberg account suggests the strong benchmark scores did not resolve every internal concern about coding performance. It does not establish that Gemini 4 performs poorly across all real-world uses. 7

Why the coding concerns matter to Google

The issue is consequential because Google is competing to build AI coding tools, an area where Bloomberg has reported concern inside the company about rivals’ progress. In April, Bloomberg reported that Google leaders were anxious about falling behind, particularly as Anthropic offered coding tools described by sources as more effective and popular with businesses. 4

A separate July report described Gemini 3.5 Pro as months behind schedule while Google worked to improve its capabilities, especially coding. Employees cited in that report worried that Anthropic and OpenAI were producing models that exceeded Gemini’s capabilities. 1 Reuters also reported that Gemini 3.5 Pro was intended to help Google catch up in AI coding tools and agentic AI tasks. 2

Those earlier reports provide context, but they concern Gemini 3.5 Pro and broader competitive worries—not proof that Gemini 4 has the same shortcomings or that Google has lost the race. Better coding performance could matter for Google’s efforts in coding tools and agentic tasks; the available sources do not establish how Gemini 4’s reported issues affect specific products or customers. 247

Was Gemini 3.5 Pro abandoned?

The available reporting supports saying that Gemini 3.5 Pro was delayed while Google worked on its capabilities, particularly coding. It does not establish that Google abandoned the model. 12

The distinction is important: a delay is not evidence that a model was canceled, and the reports about Gemini 3.5 Pro should not be treated as confirmation of Gemini 4’s performance. The Bloomberg report on Gemini 4 describes a separate set of employee concerns. 17

How Google responded

Google called it “inaccurate” to characterize Gemini 4 as underperforming in coding, according to Bloomberg’s report as relayed by ZeroHedge. That response conflicts with the anonymous employee accounts, and the sources provided do not include a public, independent test that settles the disagreement. 57

The timing also needs context: while Bloomberg described internal skepticism as Gemini 4 approached release, a same-day report said Google had unveiled Gemini 4 Argon and had not announced a timetable for general public availability. 710

How Alphabet shares reacted

After the report, Alphabet shares gave back some of their gains. They moved from more than 2% higher to about 0.5% higher near the close, according to a report citing Investing.com. That was a reduction in the day’s gains, not a decline below the previous close. 9

What the report establishes—and what it doesn’t

The reporting points to a real question for Google: whether Gemini 4’s strong benchmark performance carries over to the coding tasks employees tried. But the evidence is a mix of anonymous internal accounts and Google’s rebuttal, not a complete public evaluation. The earlier Gemini 3.5 Pro delay adds competitive context; it does not prove that the model was abandoned or that Gemini 4 is broadly behind its rivals. 157