At Google's annual developer conference on May 19, 2026, CEO Sundar Pichai asked the audience to "wait until next month." The promise was Gemini 3.5 Pro. June passed, half of July passed, and the model still had not shipped.
According to Bloomberg, the delay stems from performance shortfalls found in internal testing, particularly accuracy on code generation and complex tasks. Alphabet's stock fell more than 4% the day the report landed.
Citing anonymous sources, outlets including HackerNoon and Geeky Gadgets reported that Google scrapped a nearly finished model and restarted from pre-training. The weaknesses identified in early testing were agentic work involving repeated tool calls, mathematical reasoning, and image and graphic generation quality.
The significance is that those are not fine-tuning problems. A model's ceiling is set during pre-training — the largest and most expensive stage — and later tuning only polishes within that ceiling. If the ceiling landed too low, rebuilding is the only real option, at enormous cost.
Rival flagships shipped during the same window. OpenAI, Anthropic and Chinese open-source labs all released new models while Google passed through the period without the model it promised — exactly when developers and enterprises decide which AI to build on. Reports of senior DeepMind researchers leaving for competitors added to the picture.
Google's stopgap was to lead with cheaper models. On July 21 it released Gemini 3.6 Flash, 3.5 Flash-Lite and a security-focused 3.5 Flash Cyber, filling the flagship gap with efficient options.
Google had the familiar alternative: ship now, patch later. But it has already learned what happens when rushed AI features produce strange answers — headlines, then eroded trust. A delay costs a few days of criticism; a disappointing product gets recited for years. That pattern is covered in Shipped Friday, Pulled Saturday — AI Features Keep Launching Without Guardrails.
First, the "full rebuild" is not official. Major outlets reported the performance-driven delay, but the claim that Google discarded the model and started over rests on anonymous sourcing. On its July 22 earnings call, Google said 3.5 Pro and upgraded Flash models were "in testing with partners," and Pichai said Gemini 3.5 Pro is currently in testing while the team is already building the next generation, having begun "the most ambitious pre-training run ever" for Gemini 4.
Second, late is not the same as lost. A few months looks fatal in AI news cycles, but what gets remembered is the finished product. If the rebuilt model is strong enough, this delay will be re-read as prudence. For where the current ceiling actually sits, see OpenAI's 'Astra' Solves 10 Long-Open Math Problems — for About $2,000 in Tokens.
Throw away something nearly finished, or ship it knowing it falls short? Every product team faces that moment. Google is facing the most expensive version of it.
Bloomberg reported that internal testing showed the model falling short of targets, especially on code generation and accuracy for complex tasks.
That claim comes from anonymous sourcing in press reports, not from Google. On its earnings call the company said 3.5 Pro and upgraded Flash models are in testing with partners.
On July 21 it released Gemini 3.6 Flash, 3.5 Flash-Lite and the security-focused 3.5 Flash Cyber — cheaper, efficient models covering the flagship gap.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
Opening an AI before a search engine is becoming the default. As the path back to original creators narrows, where does competitive advantage in the AI era actually come from?
Meta expanded its vibe coding app Pocket to all U.S. users. Describe what you want and it builds a small app or game — then that output circulates in a social feed.
Meta's CEO published a 6,500-word essay on personal superintelligence that doubles as an argument against AI regulation — while Meta cut 20% of staff earlier this year.