fix: update gemini-live model endpoints and mode to realtime

[Chesars]

Relevant issues

Supersedes #18009

Pre-Submission checklist

• N/A - JSON config only, no code changes
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem

Type

🐛 Bug Fix

Changes

The gemini-live-2.5-flash-preview-native-audio-09-2025 model was incorrectly configured with mode: "chat" and REST API endpoints (/v1/chat/completions, /v1/completions), but this model only works with WebSockets (Realtime API).

What changed:

• supported_endpoints: Updated to correct realtime endpoints:
  • gemini-live-* (vertex_ai) → /vertex_ai/live
  • gemini/gemini-live-* (gemini) → /v1/realtime
• mode: Changed from "chat" to "realtime" so health checks use _realtime_health_check() (WebSocket) instead of acompletion() (REST), which would fail for this model

Files changed:

• model_prices_and_context_window.json
• litellm/model_prices_and_context_window_backup.json

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Error		Mar 4, 2026 10:44pm

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ Chesars
❌ github-actions[bot]
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

Fixes incorrect configuration for gemini-live-2.5-flash-preview-native-audio-09-2025 models, which are WebSocket-only (Realtime API) but were misconfigured with REST API settings. Changes the mode from "chat" to "realtime" so health checks use _realtime_health_check() (WebSocket) instead of acompletion() (REST), and updates supported_endpoints to the correct WebSocket routes (/vertex_ai/live for Vertex AI, /v1/realtime for Gemini API).

• Both the primary and backup JSON files are updated consistently
• The endpoint values match the registered WebSocket routes in litellm/proxy/proxy_server.py
• The mode: "realtime" value correctly maps to the realtime health check handler in litellm/litellm_core_utils/health_check_helpers.py
• No code changes required — this is a JSON config-only fix, consistent with the project's convention of storing model-specific flags in model_prices_and_context_window.json

Confidence Score: 5/5

• This PR is safe to merge — it corrects a config-only bug in JSON metadata with no code changes.
• The changes are minimal, well-scoped, and correct. Both JSON files are updated consistently, the endpoint values match existing registered WebSocket routes, and the mode value maps to a valid health check handler. This is a straightforward bug fix with no risk of regression.
• No files require special attention.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Corrects mode from "chat" to "realtime" and updates supported_endpoints to the proper WebSocket endpoints for both gemini-live model variants. Changes are consistent and match the proxy server's registered websocket routes.
litellm/model_prices_and_context_window_backup.json	Backup file mirrors the exact same changes as the primary JSON file, keeping both files in sync.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["gemini-live-2.5-flash model"] --> B{mode?}
    B -->|"Before: chat"| C["acompletion() REST call"]
    C --> D["❌ Fails — model is WebSocket-only"]
    B -->|"After: realtime"| E["_realtime_health_check() WebSocket call"]
    E --> F["✅ Correct health check"]

    G["Vertex AI variant"] --> H["/vertex_ai/live endpoint"]
    I["Gemini API variant"] --> J["/v1/realtime endpoint"]

Loading

_{Last reviewed commit: 0e1a633}