@Ed_Godshaw, your points about the 05-06 model’s performance regressions and bugs are being felt by many, and that’s valid feedback for Google on that specific checkpoint.
HOWEVER, and this cannot be stressed enough for everyone reading and for Google, the critical issue at the heart of this discussion, and the one that represents an unprecedented breach of developer trust, is the silent redirection of the gemini-2.5-pro-preview-03-25 endpoint. While the 05-06 model having its own set of problems is frustrating for sure, that’s a separate matter of model quality. the primary, urgent concern here is that a supposedly stable, dated checkpoint was changed underneath us without warning, violating established industry protocols and our fundamental expectations for API stability. This is the dangerous precedent we are focused on, and it’s vital to keep this thread centered on that specific action and the systemic damage it does to the Gemini API as a trustworthy development platform.
Google employees are typically very responsive on these forums; the noticeable silence on this specific thread for 24 hours is telling. This, especially when combined with the similar silence from @OfficialLogan on X and other platforms despite direct inquiries, hasn’t gone unnoticed by the community.
I can only speculate on the team’s rationale here. My guess is they wanted developers on the new checkpoint to do aggressive testing and iteration, but realized they had backed themselves into a corner because everyone was on a clearly dated model name.
So their options were to:
A) Deactivate that endpoint and break everyone’s API calls, or
B) Redirect the model name in an unprecedented breach of longstanding industry protocol and their own protocols, solving their internal problem, but creating a massive external problem.
They chose option B, and the fallout is clear. We’re even seeing knock-on effects like the AIDER benchmark team admitting they made a mistake in calculating the widely publicized cost figures for the original 03-25 Gemini 2.5 Pro (figures Google themselves highly promoted), but they are now unable to correct this because of the redirection issues. This kind of chaos is a direct result of Google’s decision.
Google has a choice here:
- Publicly get ahead of the backlash, admit this wasn’t ideal, and give a public commitment going forward to clarify the exact policy around dated checkpoints in
PreviewandExperimentaltiers, and rectify the current situation. - They can continue to bury this under the rug and hope it goes away, and go the way of the very recent public OpenAI blunders.
If they go with option 2, they will not have a good outcome. The trust being eroded is a valuable asset.
Even right now, a major public AI influencer and researcher, Simon Willison, co-creator of the Django Web Framework (someone Google themselves has featured in interviews with @Logan_Kilpatrick who has notably ignored emails and public calls for action on this, perhaps being directed from above), has just come out and publicly condemned Google’s actions here and their choice to break trust by redirecting the fixed dated checkpoint. I’ve been in touch with a number of AI influencers on YouTube and X, and most of them plan to run with this story and blow it up.
To any Google employees reading this: this matter needs to be escalated. A public statement, a policy clarification, and ideally, a course correction are needed ASAP. This is the first time a major AI provider has broken this longstanding implicit policy, and the precedent it sets is very concerning. We need to know if we can trust dated model identifiers from Google, or if “preview” now just means “volatile.”