feat(google-vertex): update model YAMLs [bot] - #2927
models-bot[bot] wants to merge 1 commit into
Conversation
|
/test-models |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 6d6b0f1. Configure here.
| output: | ||
| - text | ||
| mode: ocr # Google Cloud Vision OCR/text-detection has no matching #Mode enum value | ||
| mode: unknown # Google Cloud Vision OCR/text-detection has no matching #Mode enum value |
There was a problem hiding this comment.
OCR model mode set unknown
High Severity
mode and supportedModes change from ocr to unknown even though ocr is a valid #Mode value and matches this Cloud Vision text-detection model. Sibling Vertex OCR entries keep ocr. The gateway will no longer treat google/text-detector as OCR.
Additional Locations (1)
Reviewed by Cursor Bugbot for commit 6d6b0f1. Configure here.
| input_cost_per_token_batches: 5e-7 | ||
| output_cost_per_token: 0.000005 | ||
| output_cost_per_token_batches: 0.0000025 | ||
| region: global |
There was a problem hiding this comment.
Missing max_tokens limit override
Medium Severity
This file is now a fuller catalog entry with costs, and limits.max_output_tokens is 65536, above the inherited google-vertex default.yaml max_tokens cap of 32768. There is no params override, so callers stay capped at the default.
Triggered by learned rule: Override max_tokens maxValue to match limits
Reviewed by Cursor Bugbot for commit 6d6b0f1. Configure here.
Auto-generated by poc-agent for provider
google-vertex.Note
Medium Risk
Tiered pricing mode and token-cost updates directly affect billing estimates; lifecycle/status changes can hide or mislabel models in routing if consumers trust YAML without validation.
Overview
Bulk auto-sync of google-vertex model catalog YAMLs: pricing/caching, regions, lifecycle flags, and capability metadata.
Pricing & caching: Many Gemini models gain
cache_creation_input_token_cost_per_hour(and related cache read costs). Claude Sonnet 4.5 tiered pricing switches fromcumulativetomarginalacross regions. Grok 4.20 reasoning gets revised base token rates plus tiered pricing; non-reasoning context window drops from 2M to 1M. GLM-5-2 adds global costs and caps max output at 64k (down from 131k). gemini-2.5-flash-image output rates are normalized to scientific notation (values unchanged). New or filled-in costs for Lyria previews, robotics-er-2, and GLM-5-2.Lifecycle & status: Gemini 2.5 Pro retirement moves to 2026-10-20. 2.5-flash-image marked deprecated with retirement 2026-10-02. Video Intelligence models (video-speech-transcription, video-text-detection) and GLM-4.7-maas move to deprecated; Kimi K2-5 to retired. Gemini 3.1-flash-image goes preview → active with a retirement date; gemini-embedding-2 and resnet50 go active → preview. Grok 4.20 reasoning promoted to active. Palmyra X4 gets a deprecation date.
Other metadata: Claude Sonnet 4-6 adds asia-east1 pricing. 3.1-flash-image adds us/eu cost rows. 3-pro-image-preview enables prompt_caching. face-detector region
*→ global. text-detector mode/supportedModes ocr → unknown. Minor source URL updates and small limits/features tweaks (e.g. virtueguard context window, flash-lite-image PDF input).Reviewed by Cursor Bugbot for commit 6d6b0f1. Bugbot is set up for automated code reviews on this repo. Configure here.