Gemini 3.1 Pro Still Carries Google’s Pro AI Flag as 3.5 Delay Drags On

Google DeepMind’s Gemini 3.1 Pro remains the company’s flagship Pro-tier large language model in mid-2026, even though the company announced its successor more than two months ago. The model, identified in the API as gemini-3.1-pro-preview, offers a 1-million-token input context and a 64,000-token output limit, making it one of the most capable publicly available models for long-document analysis, multimodal reasoning, and agentic coding. The unusual situation stems from the delayed release of Gemini 3.5 Pro, which was unveiled as part of the Gemini 3.5 family at Google I/O on May 19, 2026, but has not yet shipped.

For enterprises and technical leaders, this matters because Gemini 3.1 Pro is not simply a placeholder. It is the current default choice for organizations building on Google’s most advanced Pro-tier capabilities, and its benchmark results place it among the strongest models available for scientific reasoning, competitive coding, and multi-step tool use. At the same time, the prolonged absence of its successor raises questions about Google DeepMind’s product cadence and whether reliability issues are holding back the next generation.

A flagship in limbo: the 3.5 Pro delay

At Google I/O on May 19, 2026, Google DeepMind announced the Gemini 3.5 family. Gemini 3.5 Flash began shipping immediately, but Gemini 3.5 Pro did not. As of July 2026, there is still no confirmed release date for the Pro-tier model. That leaves Gemini 3.1 Pro as the current flagship Pro model despite the newer Flash variant being available. The announcement was co-authored by Google DeepMind CTO Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals, and Noam Shazeer, underscoring the seniority of the team involved.

The delay has not gone unnoticed. Developer forums and industry trackers have noted that Gemini 3.5 Pro has been missing for more than two months. The frustration is compounded by the fact that competitors have continued to ship updates in the same period. Some developers speculate that Google is focusing on improving reliability and coding performance before releasing the next Pro-tier model. For organizations planning long-term AI infrastructure, the uncertainty around 3.5 Pro means that Gemini 3.1 Pro remains the most dependable current option on Google’s Pro-tier roadmap.

Benchmarks and technical profile

Gemini 3.1 Pro’s technical specifications are designed for heavy enterprise workloads. The model accepts up to 1 million tokens of input context, enough to process entire codebases, long legal documents, or extensive financial filings in a single call. It can generate up to 64,000 tokens of output. Its knowledge cutoff is January 2025, meaning the model does not natively know events after that date without retrieval or fine-tuning. API pricing is $2.00 per million input tokens and $12.00 per million output tokens for prompts up to 200,000 tokens, according to pricing data compiled by eesel.ai.

On key benchmarks, Gemini 3.1 Pro performs at or near the top of the current public leaderboard:

• GPQA Diamond (scientific knowledge): 94.3%
• Humanity’s Last Exam (academic reasoning): 44.4%
• ARC-AGI-2 (abstract reasoning): 77.1%
• SWE-Bench Verified (agentic coding): 80.6%
• LiveCodeBench Pro (competitive coding): 2887 Elo
• MCP Atlas (multi-step tool workflows): 69.2%

These numbers tell a nuanced story. The GPQA Diamond score of 94.3% is exceptionally strong for scientific question answering, while the 80.6% on SWE-Bench Verified indicates solid agentic coding ability. The LiveCodeBench Pro Elo of 2887 also places it among elite competitive coding models. The 69.2% on MCP Atlas, which measures multi-step tool workflows, suggests that the model can handle complex tool orchestration but still leaves room for improvement. For executives, these benchmarks translate into a model that can support advanced research, software engineering, and data analysis tasks, but not without supervision.

Praise, criticism, and the hallucination problem

Developers who have used Gemini 3.1 Pro in production often praise its speed and report fewer hallucinations compared with rival models such as ChatGPT. Those qualities matter in high-volume enterprise applications where latency and factual reliability directly affect user trust and operational cost.

However, the model is not immune to the failures that plague large language models. Recurring complaints include “confident hallucinations” — cases where the model states incorrect information with high certainty — and “doom loops” in agentic coding tasks, where the model repeatedly attempts the same failing approach without escaping the cycle. These failures are especially costly in agentic coding tasks, because a confident error can propagate through downstream tools and decisions before a human reviewer catches it.

The delay of Gemini 3.5 Pro has amplified these concerns. Developer frustration has grown alongside speculation that Google DeepMind is holding the next Pro model back to address exactly these reliability and coding performance issues. If that is the case, Gemini 3.1 Pro may be a transitional model that still carries known weaknesses, even as it outperforms many competitors on standard benchmarks.

Why it matters for enterprise AI strategy