I’m facing this issue as well in an agentic app. For me, it occurs when using Gemini 3 Flash; I couldn’t reproduce it with the 2.5 Flash or 2.5 Flash Lite (September preview) models.
I’ve tested both the model provided by Vertex AI and the one from Google AI Studio. The AI Studio API version always fails, returning nonsensical JSON. Vertex AI has a higher success rate, but it often includes a section starting with pasted_thought that exposes the model’s raw chain-of-thought (which is different from the thought summaries).
It started yesterday. It is interesting because if you ask it to answer with “PURE TEXT” it will do, but the answer will come inside a string. To be fair, it is unusable currently
We also stumbled across this issue using the OpenAI style endpoint for Gemini 3 models with tools. Since roughly a day, we get a Json string in the message content instead of just a normal string.
Would like to have that resolved. For now the fallback is to go back to 2.5 flash instead of 3 flash.
I’m seeing this as well on gemini flash 2.5 and 3.0. Using the HTTP API. Regression seems to have started yesterday (Dec 22). All responses are coming back as JSON now (not using structured output, setting functionCallingConfig.mode to AUTO)
the issues is still going, i removed AUTO and sometimes it sends text parts that look like tool calls without the tool name and IT NEVER produces a text part if there is a tool call, there is no reason for that, a single turn can contain both text parts and tool calls.
Did you talk to the itnernal team @Srikanta_K_N ??? ha?
Same problem here, with flash3 and gemini3 pro. I am in a refactoring process, so you can imagine that I thought the problem was with my code. I will change to Anthropic for now; I hope we get a fix on this.
Hi, we are in a refactoring process, you can imagine we thought our code was the problem, especially after I checked out the API status and it says that everything works ok.
Why don’t you update the API status? If it doesn’t work, it’s not normal to have it on green.
Hey folks, we are looking into this on our side. If anyone has a minimum reproducible code sample that would help. I have not been able to repro yet.
It does sounds like this is happening across both the Google SDK’s and the Vercel AI SDK so potentially no SDK release related but trying to track that down too. Hang tight!
