This is a summary of a Hacker News comment thread discussing the developer experience and technical aspects of the Google Gemini API, following an article (reportedly down) titled "Google Gemini has the worst LLM API". The discussion highlights confusion between the standard Gemini API (via AI Studio) and the enterprise-focused Vertex AI platform, citing issues with authentication methods (API keys vs. service accounts), JSON schema support (property ordering, structure, compatibility), multimodal input handling, documentation clarity, usage monitoring, and API reliability. While the Gemini models themselves receive positive remarks, the platform's complexity and perceived friction points for developers are frequently raised. Notably, Google Product Managers participated in the thread to acknowledge feedback, clarify points, and announce planned improvements, including a "Vertex express mode" preview and documentation revamps.
Exploring the Google Gemini API: Developer Feedback and Issues
0:00 (12 hours ago): The discussion starts with the linked article being down and an early comment highlighting an issue with property ordering in JSON schemas within the Gemini API, which can lead to unexpected output.
2:00 (10 hours ago): Multiple commenters express confusion regarding the distinction and usage of Vertex AI versus the non-Vertex APIs available through AI Studio. The non-Vertex API is noted as easier for startups/individuals (API keys, debugging tool).
2:00 (10 hours ago): Google offers an OpenAI-compatible API endpoint to simplify switching from apps using OpenAI client libraries, though some users find it missing parameters like the ability to disable "thinking".
3:00 (9 hours ago): Vertex AI is seen as Google's enterprise platform, necessary for unified cloud billing, using large cloud credits ($350K), and meeting enterprise security/compliance requirements (e.g., service accounts, data region control), despite adding complexity.
4:00 (8 hours ago): Authentication methods are a point of friction; service accounts are recommended over API keys for security but are perceived as harder to set up than simple API keys. Workload Identity Federation is suggested for CI/CD.
2:00 (10 hours ago): JSON Schema support is criticized for its limitations (e.g., lack of unions), differing format compared to other providers (OpenAI, Anthropic), and patchy support for tool calls and structured generation, causing compatibility issues with libraries that abstract over models.
3:00 (9 hours ago): The billing dashboard is described as poor, over-engineered, slow to update (up to 24h stale data), lacking clarity on model-specific costs, and missing features like hard caps on API keys.
2:00 (10 hours ago): Multimodal input handling is seen as cumbersome, requiring file uploads via the Files API or Google Cloud Storage rather than direct embedding in requests, and some client libraries have implementation issues.
2:00 (10 hours ago): The lack of an OpenAPI specification for the API, reliance on v1beta protobufs, and outdated documentation are highlighted as developer pain points.
9:00 (3 hours ago): Google Product Managers (Chris Cho, Ivan Nardini) directly engage in the comments, acknowledging the developer experience friction, thanking users for feedback, and announcing plans to address concerns, including a "Vertex express mode" preview with API keys and documentation improvements.
9:00 (3 hours ago): Specific issues mentioned include difficulty with the stop_sequences API, unclear timelines for new model releases (like Gemini 2.5 Pro in European regions), and challenges with language-specific SDKs (e.g., Java dependency resolution, retries, timeouts).
10:00 (2 hours ago): The complexity of integrating Google's OAuth flow, even for background scripts, is noted as challenging.