Google Gemini 3.5 Pro delayed 3 months as coding performance trails rivals

Google's Gemini 3.5 Pro has missed three consecutive launch deadlines after the company scrapped its original base model and restarted pretraining, leaving the search giant without a competitive flagship as rivals ship production-ready alternatives.

"The delay is about agent work, not chatbot polish," Dave Barr, a technology analyst at Startup Fortune, said. "Enterprise buyers now test models on coding, tool use, and long-horizon tasks — and Google's internal evaluation found the original model couldn't clear those bars."

The rebuilt model targets a 2 million token context window — double Gemini 2.5 Pro's 1 million cap — and a new Deep Think reasoning mode for multi-step logic, according to leaked specifications reported by multiple outlets. But the original version showed structural failures in recursive tool-calling and complex SVG generation, according to HackerNoon, citing unnamed internal sources. Gemini 3.5 Flash, which launched May 19, already outperforms Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2% vs. 70.3%) and MCP Atlas (83.6% vs. 78.2%), per Google's official launch post — making the Pro model's regression on those same tasks a fundamental capability gap rather than a finishing problem.

Alphabet shares slipped as much as 3.2% on the initial delay report, and the stock fell roughly 5% to 7% in late June after key researchers — including Transformer co-author Noam Shazeer and Nobel laureate John Jumper — departed for OpenAI and Anthropic, wiping an estimated $225 billion in market capitalization. Google trades at roughly 22 times forward earnings, a discount to Microsoft's 30 times, reflecting the market's growing skepticism about its AI monetization timeline.

The Rebuild Decision Signals a Structural Problem

Pre-training is the most expensive phase of building a frontier AI model — a months-long run on vast datasets that establishes the model's fundamental capability ceiling. Fine-tuning and reinforcement learning can refine within those bounds but cannot raise them. When Google chose to restart pre-training rather than continue refining, it was conceding that the original model's ceiling was in the wrong place.

According to reporting from Geeky Gadgets, citing World of AI, the rebuilt model has now encountered a third delay due to frequent hallucinations and inconsistent outputs in real-world workflows — distinct failure modes from the recursive tool-calling problems that drove the first rebuild. Google DeepMind has registered model names including Gemini 3.6 Flash and Gemini 3.5 Flash Light, suggesting the company is preparing stopgap releases while Pro development continues.

The talent exodus compounds the technical challenges. Shazeer — whom Google spent a reported $2.7 billion to bring back from Character.AI in 2024 — left for OpenAI in June. Jumper joined Anthropic the same week, along with Google researchers Jonas Adler and Alexander Pritzel. DeepMind CEO Demis Hassabis said at Cannes that Google has "by far the biggest and broadest research bench of any of the labs out there," but headcount alone does not determine who ships first.

Rivals Already in Production With Competitive Pricing

The competitive landscape has hardened considerably since Google I/O in May. OpenAI's GPT-5.6 reached general availability July 9 across three tiers — Sol at $5 per million input tokens and $30 per million output tokens, Terra at $2.50 and $15, and Luna at $1 and $6 — all sharing a 1.05 million token context window. OpenAI confirmed Sol sets a new state of the art on Terminal-Bench 2.1, per its July 9 launch announcement.

Anthropic's Claude Fable 5 launched July 1 after a brief export-control review, with usage-based fees attached to its most capable model. Grok 4.5 from SpaceXAI launched July 8 at $2 per million input tokens and $6 per million output tokens, co-trained with developer session data from Cursor, the coding editor SpaceX agreed to acquire in June. Independent benchmarking organization Artificial Analysis placed Grok 4.5 fourth on its Intelligence Index at launch.

That leaves Google as the only major frontier lab without a 2026 flagship in general production on the date its own CEO promised delivery. Prediction markets on Polymarket now show 81% probability of a July 31 delivery as the leading outcome, with a separate market showing "August 7" at 73% — reflecting traders' growing confidence that neither July 17 nor July 24 will produce a launch.

What Developers Should Watch Next

For teams that built plans around a July 17 Gemini 3.5 Pro launch, the watch signal remains the official one: gemini-3.5-pro appearing as a generally available model in the public Gemini API documentation. Model name registrations, leaks, and unnamed internal sources do not constitute a launch.

One deadline developers cannot ignore is July 24, when DeepSeek's legacy API aliases — deepseek-chat and deepseek-reasoner — stop responding, with no announced extension. Teams calling either alias need to migrate to deepseek-v4-pro or deepseek-v4-flash before that date. DeepSeek V4 offers a 1 million token context window and competitive pricing, but teams should note that DeepSeek is headquartered in China and subject to the country's National Intelligence Law, which requires all organizations to cooperate with national intelligence work on demand.

For workloads that fit within a 1 million token context window, Gemini 3.5 Flash, GPT-5.6 Terra, or Claude Fable 5 represent production options available today. Google can still make this quarter work if the rebuilt Pro model ships with verified benchmark improvements and competitive pricing. But the reported $15 per million input tokens and $60 per million output tokens — if accurate — would price Gemini 3.5 Pro above all current production competitors, requiring strong benchmark differentiation to justify. The rebuild decision tells developers that Google's internal evaluation concluded the competitive gap was real enough to justify a costly intervention. Whether that intervention produced a model that closes the gap against GPT-5.6 Sol and Claude Fable 5 is a question only the official model card and independent benchmarking can answer.

This article is for informational purposes only and does not constitute investment advice.