Google finds itself once again explaining why its flagship AI model isn’t ready. Reports surfaced Thursday that Gemini 3.5 Pro missed yet another launch target. The reason? Coding performance that still lags behind what executives demanded internally.
This isn’t the first stumble. The Next Web detailed similar issues drawing from a Bloomberg investigation. Ten current and former employees described mounting frustration. Anthropic and OpenAI keep pulling ahead. Senior researchers continue to depart for rival labs.
But the story runs deeper than one missed deadline. It reveals structural challenges at a company long known for moonshot ambitions yet now wrestling with execution in one of tech’s hottest arenas. Alphabet shares dropped more than 4% on the news. Investors sense the competitive gap widening.
The original plan looked straightforward. At Google I/O in May, CEO Sundar Pichai pointed to June for general availability of the new Pro model. That date passed quietly. A July 17 target slipped too. Sources tell Bloomberg the team even rebuilt parts of the model last month with fresh training data focused on programming tasks. Results disappointed.
Internal benchmarks told a harsh story.
Hallucinations persisted in complex scenarios. The model struggled with recursive tool calling across long chains of decisions. Structural consistency broke down when generating intricate visual layouts or multi-file code modifications. These aren’t edge cases. They define what modern agentic coding systems must handle.
Google’s own Gemini 3.5 Flash, released in May, picked up some slack. It posted solid numbers: 76.2% on Terminal-Bench 2.1 and 83.6% on MCP Atlas, according to the company’s official announcement. The smaller model runs four times faster than some frontier competitors. Yet Flash serves as a stopgap. Enterprise customers want the heavier Pro version for sustained, high-stakes work.
Comparisons sting. OpenAI’s GPT-5.6 family and xAI’s Grok 4.5 landed recently with stronger results in similar tests. A TechCrunch report from February highlighted Gemini 3.1 Pro setting records on several benchmarks including SWE-Bench. Progress arrived. Then expectations rose faster.
Developers noticed inconsistencies too. Forums lit up with complaints. One Cursor community thread described performance swinging wildly day to day. “Great yesterday, terrible today,” users posted. Context handling weakened. Outputs grew simplistic. Timeouts increased. Similar gripes appeared on Google support pages and Reddit.
And the internal dynamics complicate matters further. DeepMind, Google Cloud, Android, and consumer teams all chase their own AI coding initiatives. Sergey Brin pushed for speed on these tools. Yet competing factions slowed decisions. Some engineers still insist critical code belongs in human hands to meet quality bars. That tension hasn’t vanished.
Efforts to consolidate gained traction. Chief AI Architect Koray Kavukcuoglu works to unify tooling. A new DeepMind team under research engineer Sebastian Borgeaud targets coding specifically. At a recent Cloud conference, executives boasted that 75% of Google’s own code now comes from AI systems. Much of the developer workflow funnels through Antigravity, an internal platform handling data, memory, and safety.
Those claims sound impressive. Reality for external users varies. Figma product manager Rodrigo Davies praised an earlier Flash version for striking “a sweet spot of speed and quality” in their design assistant. Contrast that with Platzi CEO Freddy Vega. His team ditched the model after finding it more expensive, slower, and less capable than predecessors. They switched to Anthropic.
The talent exodus adds pressure. Multiple reports confirm senior Gemini researchers joined Anthropic. Frustration with Google’s position in the race drove some exits. Compute constraints inside the company don’t help. Teams compete for resources. Access to rival models like Claude stays restricted to select groups.
Recent independent evaluations paint a mixed picture. DeepMind’s own published benchmarks for Gemini 3.1 Pro show leadership in areas like LiveCodeBench Pro with an Elo of 2887 and 80.6% on SWE-Bench Verified. Yet newer agentic tests expose gaps. A TechTimes article published just yesterday cited internal sources saying the rebuilt 3.5 Pro couldn’t clear reliability thresholds for complex SVG generation or extended tool chains.
Google isn’t standing still. The company launched Gemini CLI, an open-source terminal tool, and made Code Assist free for individual developers with generous usage limits. Antigravity, the agent-first coding interface tied to Gemini 3, offers multi-pane editing with browser control. These moves aim to keep developers engaged while the flagship model catches up.
But the delay hits at a pivotal moment. Rivals ship updates monthly. OpenAI, Anthropic, and even xAI set aggressive paces. Google’s measured approach emphasizes safety and quality. Spokespeople repeat that the team ships “quickly across a wide range of models” and tests upgrades with partners. They point to government discussions on testing standards.
Still, the pattern worries observers. This marks the third missed target for 3.5 Pro. June became July. July became uncertain. No firm new date exists. Enterprise roadmaps bend accordingly. Developers build around Flash for now, watching benchmarks and waiting for the Pro upgrade that promises deeper reasoning and fewer errors.
The stakes extend beyond bragging rights. Coding agents represent the next frontier. They don’t just autocomplete lines. They plan features, debug across repositories, orchestrate tool calls, and verify outputs. Companies bet workflows on these systems. Reliability matters. Consistency matters more.
Google built its empire on search and infrastructure. AI demands different muscles: rapid iteration, decisive prioritization, and tolerance for public experimentation. The company’s size brings advantages in data and compute. It also breeds the very factions now slowing progress.
Whether the next rebuild succeeds remains unknown. Training runs continue. Feedback loops tighten. Yet each delay hands rivals more time to widen their lead. Users grow impatient. Talent scans the exits. And Wall Street watches the stock react.
One thing feels clear. The AI coding race won’t pause for Google to sort its house. The models improve relentlessly. So do expectations. The company that cracks sustained agentic performance at scale will claim significant ground. For now, Google fights to prove it still belongs in that conversation.
WebProNews is an iEntry Publication