openclaw - 💡(How to fix) Fix [Bug]: google/gemini-3.1-flash-lite silently rewritten to deprecated preview model — will break after May 25 deprecation deadline [1 comments, 2 participants]

Bug type

Regression (worked before, now fails)

Beta release blocker

No

Summary

When google/gemini-3.1-flash-lite (GA) is set in config or used directly, OpenClaw silently rewrites the model ID to google/gemini-3.1-flash-lite-preview (deprecated). Google has announced the preview variant will be shut down on May 25, 2026. After that date, any OpenClaw install using flash-lite will break silently with no clear error.

Steps to reproduce

openclaw infer model run --model google/gemini-3.1-flash-lite --prompt "hi" Expected: model: gemini-3.1-flash-lite Actual: model: gemini-3.1-flash-lite-preview

Confirmed on v2026.4.23. The GA model ID is silently rewritten to the deprecated preview variant at the transport layer.

Expected behavior

openclaw infer model run --model google/gemini-3.1-flash-lite --prompt "hi" Should use the GA model gemini-3.1-flash-lite and return a response with: model: gemini-3.1-flash-lite

Actual behavior

The model ID is silently rewritten to the deprecated preview variant before the API call is made. The response shows: model: gemini-3.1-flash-lite-preview No warning or error is surfaced. After Google's May 25 deprecation deadline, this will cause silent failures for all users relying on gemini-3.1-flash-lite as primary or fallback model.

OpenClaw version

v2026.4.23 Also confirmed present in v2026.5.18 (latest stable) per other reporters in #80380.

Operating system

Rocky Linux

Install method

No response

Model

google/gemini-3.1-flash-lite

Provider / routing chain

OpenClaw → Google AI Studio (GEMINI_API_KEY) → gemini-3.1-flash-lite (silently rewritten to gemini-3.1-flash-lite-preview)

Additional provider/model setup details

No response

Logs, screenshots, and evidence

$ openclaw infer model run --model google/gemini-3.1-flash-lite --prompt "hi"

provider: google
model: gemini-3.1-flash-lite-preview

Tested on v2026.4.23. The provider field confirms Google AI Studio is reached, but the model field shows the GA model ID has been rewritten to the deprecated preview variant before the API call.Tested on v2026.4.23. The provider field confirms Google AI Studio is reached, but the model field shows the GA model ID has been silently rewritten to the deprecated preview variant before the API call. Also confirmed present in v2026.5.18 (latest stable) per other reporters in #80380.
The rewrite logic is likely in the model ID normalization path (possibly extensions/google/provider-registration.ts or similar). Searching for flash-lite-preview in the Google provider source should find it quickly.
Note: the changelog for recent betas (5.10–5.16) shows repeated commits normalizing retired Gemini Pro Preview model IDs across multiple files — suggesting the rewrite logic is not centralized but scattered across the codebase. The same pattern likely applies to gemini-3.1-flash-lite-preview. A grep for flash-lite-preview across the full source (not just provider registration) may be needed to catch all rewrite paths.

Confirmed on v2026.4.23: running `openclaw infer model run --model google/gemini-3.1-flash-lite --prompt "hi"` shows `model: gemini-3.1-flash-lite-preview` in the output — the GA model ID is silently rewritten to the preview variant.

The rewrite logic is likely in the model ID normalization path (possibly extensions/google/provider-registration.ts or similar). Searching for flash-lite-preview in the Google provider source should find it quickly.

Note: the changelog for recent betas (5.10–5.16) shows repeated commits normalizing retired Gemini Pro Preview model IDs across multiple files — suggesting the rewrite logic is not centralized but scattered across the codebase. The same pattern likely applies to gemini-3.1-flash-lite-preview. A grep for flash-lite-preview across the full source (not just provider registration) may be needed to catch all rewrite paths.

Also confirmed present in v2026.5.18 (latest stable) per other reporters in #80380.

Impact and severity

All users relying on gemini-3.1-flash-lite as primary or fallback model will silently break after May 25. No error is surfaced — the model just stops working.

Additional information

From changelog analysis across 5.10–5.16 beta series, the rewrite logic for retired Gemini model IDs is not centralized — it is scattered across multiple files in the codebase (repeated commits normalizing the same IDs across different paths confirm this). A grep for flash-lite-preview across the full source tree (not just extensions/google/provider-registration.ts) is likely needed to catch all rewrite paths. This is consistent with why a local Codex fix attempt also failed.