We wanted to give folks a heads up that Gemini 2.5 Pro Preview 05-06 is scheduled for deprecation on June 19, 2025, given that we just launched an updated preview version Gemini 2.5 Pro Preview 06-05 (Gemini 2.5 Pro Preview in AIS) earlier this week, and will be GAing in a matter of weeks.
We recommend switching to the updated model which shows improvements across the board on key benchmarks for coding, science, multimodal understanding and reasoning. We have also improved creative writing, style and structure based on feedback from our previous releases.
Please let us know if you have any issues or concerns with switching to the updated Pro model by June 19, 2025.
These benchmarks are misleading, inflated! Claude Sonnet 4 (Extended Thinking) consistently corrects content generated by Gemini 2.5 Pro Preview 06-05 and it acknowledges its own limitations and reasoning gaps; however, the benchmark results indicate otherwise.
Providers should stop publishing benchmark results and allow the quality and standards of the models to speak for themselves. It is the quality that determines the bench results, not the other way around. I presume that you do not extensively use your own models.
The only advantage your models have is their extensive token count, which is unparalleled.
This news is quite disappointing. After conducting actual testing, our team found that the 06-05 version performs significantly worse than the 05-06 version, yet now the official team is planning to discontinue the better-performing 05-06 version.
It’s frustrating to see that the version with superior performance in logical reasoning, text comprehension, and generation quality (05-06) is being phased out in favor of an inferior one (06-05). This seems like a step backward for users who rely on consistent, high-quality performance.
I hope the official team reconsiders this decision or at least addresses the performance issues in 06-05 before making the switch mandatory. Discontinuing the better version doesn’t serve users’ best interests.
Glad to see that the 0605 performance issue is verified! I had long been suspecting its performance compared with 0506 and got very frustrated because it contradicts benchmarks ranking.
I notice that 0605 tends to ‘think’ (i.e. more liberty in generation process), but its accuracy and comprehension is objectively missing!
I was able to simulate thinking process in the answer itself. I noticed that 06-05 is extremely lazy and it’s trying to avoid long analysis as much as possible
It feels that laziness mostly in the decision making part. For example, if you start deep analysis by 05-06 and 06-05 would continue that, everything is completely fine. But if you ask to do the same 06-05, it would cut corners as much as possible.