Three missed deadlines in two months

The story began on May 19 at the Google I/O 2026 developer conference. Google CEO Sundar Pichai introduced the faster Gemini 3.5 Flash model there and, regarding the more powerful Pro version, uttered a sentence that's now coming back to haunt him: "Give us a month, you'll get it." June ended without a release. July 17 became the unofficial backup date — and according to current information, it won't make that one either.

According to a Bloomberg report from July 16, Google is "months behind" on releasing Gemini 3.5 Pro as it tries to improve its capabilities, especially in programming. At the end of June, Google reportedly updated the training data in an effort to improve the model's coding skills, but "the results were disappointing." 9to5Google adds that this suggests a de facto restart of development between the May conference and the missed deadline.

Google officially commented on the situation only in general terms: "We are currently testing 3.5 Pro, an improved Flash model, and other models with partners. We are rapidly delivering a wide range of models and keeping them cost-effective for customers."

Where exactly the model is failing

Unofficial reports from recent weeks paint a more detailed picture. According to Tech Times, the original model had two structural problems: it couldn't maintain consistency during recursive tool calling — that is, multi-step chains where an AI agent calls one tool, derives the next step from the result, and continues across dozens of decisions. This capability is the foundation of agentic programming, the very discipline Google wants to dominate with the 3.5 generation.

Google DeepMind therefore reportedly took a radical step: it scrapped the nearly finished base model and launched a new pre-training from scratch on the Gemini 3 foundations. Pre-training is the most expensive phase of AI model development — it sets the ceiling for the model's capabilities, which subsequent fine-tuning cannot raise. But even the rebuilt model, according to the latest information, didn't pass: it's plagued by frequent hallucinations (fabricating false information) and inconsistent performance in real-world tasks, where it falls short of OpenAI's GPT-5.6.

The Flash paradox: the cheaper model outgrew the expensive one

The delay also has a curious internal dimension. The cheaper Gemini 3.5 Flash, which Google released back in May, beats the older flagship Gemini 3.1 Pro in key benchmarks: on Terminal-Bench 2.1 it scores 76.2% versus 70.3%, and on MCP Atlas 83.6% versus 78.2% — at a fraction of the price ($1.50 per million input tokens). The new Pro version therefore has to offer significantly more to justify its premium price. A model only slightly better than its own cheaper variant wouldn't stand a chance in the market.

The competition isn't sleeping

The timing is extremely awkward for Google. Anthropic released Claude Fable 5 at the beginning of July, OpenAI made GPT-5.6 available on July 9 in three tiers (the most powerful, Sol, costs $5 per million input tokens), and Grok 4.5 from xAI arrived on July 8 with an aggressive price of $2 per million tokens. Google is thus the only major AI lab whose flagship model for this year is still missing — despite the deadline being publicly promised by the CEO himself.

The situation is worsened by the June wave of top researcher departures. Transformer architecture co-author Noam Shazeer left for OpenAI, Nobel laureate John Jumper (AlphaFold) and two other researchers headed to Anthropic. Markets wiped an estimated $225 billion from Alphabet's value at the time. According to CNBC, Alphabet shares also fell following Wednesday's report of another delay.

An interesting detail comes from prediction betting markets: on the Polymarket platform, traders assign the highest probability to a Gemini 3.5 Pro release in early August. Meanwhile, Google has registered the model names Gemini 3.6 Flash and Gemini 3.5 Flash Light — suggesting it may fill the gap with faster but less ambitious models.

What this means for Czech users and developers

For regular users of the Gemini app, which is fully available in Czech, practically nothing changes — it continues to run on the Gemini 3.5 Flash model, which handles most everyday tasks, and the Google AI Pro subscription works without changes. Czech developers who were counting on deploying Gemini 3.5 Pro in production applications, however, should not treat the model as a firm dependency. Until gemini-3.5-pro appears in the official API documentation, it's just promises — and for demanding programming tasks, realistic choices today are GPT-5.6, Claude Fable 5, or Gemini 3.5 Flash itself.

The whole episode shows just how brutally the pace of the AI race has accelerated. A year ago, a two-month model delay wouldn't have excited anyone. Today, when — according to April data — 75% of new code at Google is written with AI assistance and companies are standardizing their development tools in real time, every missed month can mean customers who never come back. Google still has the largest research team in the field, and distribution through Android, Search, and Workspace gives it an advantage the competition doesn't have. The question, though, is whether the outcome of the race is decided by team size — or by who actually delivers.