Gemini 3.5 Pro launch slips after coding benchmark failures

Google’s Gemini 3.5 Pro is reportedly months behind schedule after its coding capabilities failed to meet internal expectations, according to a Bloomberg investigation published on Thursday 16 July. The report, by Julia Love and Davey Alba, cites people familiar with the matter and says the setbacks have frustrated engineers, researchers and managers at the company.

The unlaunched model was presented as Google’s next flagship AI system at its I/O developer conference in May, when chief executive Sundar Pichai said it would arrive the following month. That deadline passed without a public release, and there is still no confirmed general-availability date as of Friday 17 July.

Alphabet shares closed down 4.4% on Thursday, wiping $200 billion from the company’s market capitalisation, following Bloomberg’s report. Alphabet is due to publish its second-quarter results on Wednesday 22 July.

Coding work reportedly fell short

Bloomberg says Gemini 3.5 Pro’s software-development abilities have been a particular problem. The model was expected to be Google’s most powerful offering, but its coding performance was reportedly below the company’s own targets at a moment when OpenAI and Meta have released newer models that are said to outperform Google’s current systems at generating code.

Google reportedly updated the training data used for Gemini in late June in an attempt to improve those skills. The results were disappointing, Bloomberg said, suggesting a significant reset in development between the I/O announcement and the missed June launch window.

The problem is not solely technical. Bloomberg reported that Google’s efforts to move quickly in AI coding have been slowed by competing internal groups. Google Cloud, Google DeepMind and the Android team are all developing coding tools, with some consumer-product teams also involved. Sergey Brin and others were said to be pushing for quicker progress, but the number of stakeholders across Google’s vast product line can slow a model’s route to release.

Some engineers have also resisted the broader use of AI-written code, Bloomberg reported, arguing that important code should remain human-written to meet Google standards. Early users of Gemini reportedly faced restrictions on using it to write or analyse software because of concerns that proprietary code could find its way into AI training data.

Even staff expected to use AI for coding have encountered limits on access to computing capacity, according to Bloomberg, because of competition for processing power within Google.

Google says testing continues

Google did not confirm a release date or directly address Bloomberg’s account of the coding shortfall. A spokesperson said: “We're currently testing 3.5 Pro, an upgraded Flash model, and other models with partners, and we're productively engaged with the U.S. government.”

The company added: “We're shipping quickly across a wide range of models while keeping them highly cost-effective for customers.”

Google has not published a model card, system card, pricing or public benchmark results for Gemini 3.5 Pro. Reports have suggested that the model could support text and image inputs, include a Deep Think reasoning mode and offer a two-million-token context window, but Google has not confirmed those specifications.

As previously reported, a 17 July launch had been tipped for Gemini 3.5 Pro, alongside the reported two-million-token window. That date has now arrived without a public release, reinforcing that the earlier timing was not a confirmed Google announcement.

Pressure grows in the AI race

Bloomberg’s report says the delay has worried many within Google, with 10 current and former employees describing concerns that the company could lose ground to Anthropic and OpenAI. The talent situation has added to those worries: four senior DeepMind researchers reportedly left for rivals in a single week, including Gemini co-lead Noam Shazeer, who joined OpenAI, and John Jumper, Jonas Adler and Alexander Pritzel, who went to Anthropic.

Google is attempting to consolidate its AI coding work. Chief AI Architect Koray Kavukcuoglu is working with the main engineering organisation to unite internal coding tools, while DeepMind formed a dedicated AI-coding team earlier this year led by research engineer Sebastian Borgeaud.

The company has already made AI-generated code a substantial part of its own development process. Bloomberg reported that, as of April, 75% of new code at Google was generated by AI and approved by engineers, up from 50% last autumn.

Flash remains Google’s live offering

For now, Gemini 3.5 Flash is the widely available newer model. It became the default across the Gemini app and Search’s AI Mode, and Google says the app serves more than 900 million monthly users in 230 countries, with daily queries up sevenfold.

But the reception has not been uniformly positive. Freddy Vega, chief executive and founder of education platform Platzi, told Bloomberg that Gemini 3.5 Flash is more expensive and slower than Google’s earlier 3.1 Flash model, while remaining less capable than premium competitors. He said its structured-data performance could be unreliable, and that his team had shifted some work to Anthropic’s Claude 3.5 Sonnet.

For ordinary users, a delayed Pro model does not immediately remove the Gemini tools already built into Google’s services. The bigger issue is whether Google can deliver a genuinely stronger premium model for complex reasoning and coding before rivals further establish themselves as the default choice for those demanding tasks.