Google Announces Gemini 4 Pre-Training While 3.5 Pro Remains Stuck in Development

Google is playing a familiar tune: announce a future flagship model while the current one remains stuck in development hell.

During Alphabet's Q2 2026 earnings call on Tuesday, CEO Sundar Pichai confirmed that Gemini 4 has officially entered pre-training. Calling it "our most ambitious training run yet," Pichai positioned the model as essential for Google's future competitiveness. "We will need Gemini 4 to compete at the next frontier," he stated, specifically citing coding and agentic coding capabilities as areas where the company needs to improve.

The announcement landed just one day after Google released three new models in the Flash family: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber. But buried in that press release was a quieter admission that Gemini 3.5 Pro, the frontier model originally promised for June 2026, remains in "partner testing" with no public release date in sight.

The Missing Flagship

Google's 3.5 Pro situation has become something of a running joke in AI circles. The model was first announced at Google I/O in mid-May 2026, with executives promising a June release. That deadline came and went. A revised July target was also missed. Now, according to reports from Bloomberg and multiple industry sources, the model has seen at least three separate delays.

The official explanation centers on coding performance. During closed testing, Gemini 3.5 Pro reportedly failed to meet internal benchmarks for code generation. A late-June update to the training data, intended to address these shortcomings, apparently did not deliver the expected improvements.

This leaves Google in an awkward position. While OpenAI and Anthropic have both shipped new frontier models in recent weeks, Google is the only major lab without a 2026 flagship in general production. As one analyst noted, the delays have shifted perception of Google from "leading edge to trailing edge."

Rivals Take Notice

Competitors have not been shy about poking the bear. When Google announced Gemini 4 pre-training, OpenAI's Thibault Sottiaux replied on X with a pointed comment: "Hope it finishes one day too!" Meta's chief AI officer Alexandr Wang offered a more succinct jab, posting "gemini who?" in response to benchmark comparisons.

The mockery stings because it touches on a real vulnerability. Google's AI research has long been respected, but translating research into competitive products has been a recurring challenge. The Gemini 3.5 Pro delays are the latest example of a pattern where promising announcements fail to materialize on schedule.

What Google Actually Shipped

To its credit, Google has not been idle. The three Flash models released this week represent meaningful improvements to the workhorse tier of the Gemini family.

Gemini 3.6 Flash takes over as the new default workhorse model. It costs $1.50 per million input tokens and $7.50 per million output tokens, making it cheaper than the previous generation while improving performance. On the DeepSWE coding benchmark, scores jumped from 37% to 49%. MLE Bench scores rose from 49.7% to 63.9%. Token efficiency improved by roughly 17% overall, with Google claiming up to 65% efficiency gains on specific tasks.

Gemini 3.5 Flash-Lite targets high-throughput applications with a focus on speed. It delivers 350 output tokens per second while costing just $0.30 per million input tokens and $2.50 per million output tokens. For applications where latency matters more than peak capability, it represents a compelling option.

Gemini 3.5 Flash Cyber is a specialized model for cybersecurity applications, integrated into Google's CodeMender code security agent. On the CyberGym benchmark, it scores 83.2%, within striking distance of OpenAI's GPT-5.5-Cyber at 85.6%.

These are solid releases. They just are not the frontier model that Google promised and that the market expected.

Brain Drain Adds Pressure

The technical challenges come alongside human ones. In June 2026, two prominent Google AI researchers departed: Noam Shazeer, a co-lead on Gemini, left for OpenAI, while John Jumper of AlphaFold fame joined Anthropic. Both moves reportedly stemmed from concerns within DeepMind about Google's position in AI coding tools.

The departures highlight a risk for Google. As AI talent becomes more mobile and startup funding remains plentiful, keeping top researchers engaged becomes harder when competitors ship faster and capture more attention.

Looking Ahead

Pichai's comments suggest Google knows exactly where it stands. The company is betting that efficiency and cost-effectiveness matter as much as raw capability. With token costs becoming a significant factor in AI deployment, Google's focus on cheaper, faster models could pay off even if it lacks the absolute best model on the market.

Still, the gap between promises and delivery creates a narrative problem. Announcing Gemini 4 pre-training before shipping Gemini 3.5 Pro reads as either confidence in the pipeline or an admission that the current generation is not worth waiting for. The market seems unsure which interpretation is correct.

For developers and enterprises, the practical takeaway is clear: Google has capable models available today, but if you need cutting-edge frontier performance, you will need to look elsewhere. At least until Gemini 4 actually ships. Whenever that might be.