At I/O in mid-May, Google announced Gemini 3.5 Flash and told the world the Pro version — the flagship — would arrive in June. June came and went. According to Bloomberg, Gemini 3.5 Pro is now months behind schedule, and the reason is specific: the model isn’t good enough at coding, and it falls short on the complex, long-horizon reasoning that agentic AI runs on.
The fix Google tried didn’t work. In late June, engineers updated the model’s training data specifically to lift its coding skills — the results, per Bloomberg’s reporting, were disappointing. Tech Times reports the rebuilt model has now missed three internal launch deadlines, with reliability problems including frequent hallucinations and benchmark results short of OpenAI’s GPT-5.6. Google’s official line — it is “currently testing 3.5 Pro, an upgraded Flash model, and other models with partners” — reads like a plan to ship the mid-tier stopgap first while the flagship stays in the shop.
The timing stings. In the past week alone, Mira Murati’s Thinking Machines shipped a 975-billion-parameter open-weights model, Microsoft readied a security agent assembled from its rivals’ models, and Anthropic’s free Fable 5 window turned into a live referendum on what frontier AI is worth. Meanwhile, ten current and former employees described the mood inside Google to Bloomberg as frustration — a fear the company is losing ground while Anthropic and OpenAI ship models that outperform Gemini.
Why coding is the wall
Coding stopped being one benchmark among many; it’s the product. Developers are how a lab wins distribution — they build the agents, integrations, and internal tools that lock an enterprise into one stack. Long-horizon reasoning is the other half of the same sale: it’s what separates a chatbot from an agent that can plan, execute, and self-correct across a multi-step job. Miss on both and you haven’t missed a spec sheet — you’ve missed the two capabilities enterprises are actually budgeting for in 2026.
Our take: The story isn’t that Google is late — it’s what the lateness reveals. A training-data refresh used to buy visible gains; at the frontier it now buys “disappointing.” The gap between announcing a model at a keynote and clearing internal evals is where the AI race actually lives now, and shipping the mid-tier model while the flagship slips is becoming the standard play across the industry. If you build on these platforms, the lesson is blunt: build on what’s generally available, never on a roadmap date. Keynote promises are currently the least reliable spec in tech.
What to watch
- The Flash stopgap. Google says an upgraded Flash model is in partner testing — if it ships within weeks while Pro stays dark, the two-tier strategy is confirmed.
- Alphabet earnings Wednesday. Analysts will ask about the delay directly — it lands in the middle of the week the entire AI trade gets audited.
- Benchmarks on arrival. Whenever 3.5 Pro lands, the number that matters is the gap to GPT-5.6 on coding — the target it reportedly keeps missing.