Gemini API

Google removed free tier access to Gemini 2.5-Pro entirely and slashed Gemini 2.5-Flash daily limits by 92% (250 to 20 requests) with no advance notice, email alerts, or grace period. Production applications broke overnight with 429 quota exceeded errors.

Gemini API enforces unexplained geographic restrictions preventing developers in certain regions from even requesting API keys. This creates impossible barriers for multinational enterprises, where half of development teams cannot access the API while the other half are stuck in approval queues.

Rate limits for Gemini API are managed in Google AI Studio rather than the Google Cloud Console, creating a split configuration burden. API keys generated in Cloud default to restrictive Tier 1 limits, requiring manual re-import and upgrade requests that take days and cause production downtime.

The Gemini API fails randomly approximately 1% of the time and returned errors 30% of the time in some cases, requiring mandatory retry logic. Response times are highly variable (30 seconds to 4 minutes for identical queries), making it unsuitable for production features with high uptime guarantees without multi-provider failover.

Gemini API onboarding requires 30-45 minutes of mandatory tasks including GCP project creation, Cloud Console navigation, billing setup, and government ID upload in PNG format. Individual developers must navigate enterprise infrastructure designed for organizations, violating DX best practices and causing developer abandonment before technical evaluation.

Developers face indefinite approval delays (weeks or longer) for API key requests with opaque rejection messages providing no actionable feedback. The approval process lacks status updates, timelines, or clear requirements, causing developers to abandon Gemini for OpenAI or Anthropic.

When Google rolls out new model versions, previously stable models like Gemini 1.5 Pro and 2.0 Flash experience intermittent failures and massive latency spikes (milliseconds to 15+ seconds), with issues persisting for days. Function-calling feature has intermittently failed for 3+ day periods.

Gemini API produces highly variable outputs for identical or similar prompts, making it unreliable for production use. Same prompt may generate well-structured output one moment and completely disorganized output the next, breaking workflows that require predictable results.

Gemini API experimental models have dynamic rate limits that adjust without clear communication. Multiple instances of quota reductions occurring suddenly (August and December 2025) with yo-yo patterns, creating unpredictable constraints for production applications.

Google AI Studio API has been unreliable for extended periods (2+ weeks) while the status page reports normal operation. OpenRouter reliability graphs show significant problems especially on Pro model, creating false confidence for developers.

Developers experience unexplained access rejections (e.g., "not allowed" to use Gemini Pro with CLI) despite having valid API keys and paying for the service. No reason is given and there is no documented recourse, creating frustration and blocking workflows.

Gemini API exhibits 'overthinking' behavior where it makes numerous unnecessary tool calls to accomplish simple tasks, causing unexpected cost spikes. One user reported $1 per minute in charges from only 18 API calls due to the model's inability to efficiently execute simple operations.

Gemini API enforces a hard cap of 1000 requests per hour, which is insufficient for production-scale applications. Solo developers can manage, but scaling immediately hits walls triggering '429 Too Many Requests' errors.

Developers forced to switch from Gemini 2.5-Pro to Flash due to free tier removal experience noticeable quality loss. Complex reasoning, code generation, and nuanced analysis all degrade, requiring complete prompt restructuring to maintain acceptable output.

Gemini API has undocumented or poorly documented default settings that cause problems: maxOutputTokens defaults to 8K (truncating long outputs), temperature is locked at 1.0, and TTFT can reach 29 seconds. Developers must manually discover and override these 'factory settings' or face broken functionality.

Google offers three separate APIs (Gemini API, Vertex API, and TTS API) with different stability levels and missing features across each. These APIs have separate keys and billing setups, creating integration complexity and forcing developers to choose between prototyping-only solutions and production alternatives.

Gemini API pricing starts at $99/month for basic features with additional per-request costs that scale steeply with usage. For solo developers and small teams, production-scale usage becomes financially unviable. Competing APIs offer better value propositions.

The Gemini API documentation is fragmented across multiple pages with inconsistent examples, missing response body examples, and undocumented requirements (e.g., alt=sse parameter). Tools documentation lacks practical examples, and unified guidance is absent.

Gemini API has a hard cap on input length that truncates important data from the end of requests. In testing with 80 customer feedback forms, the API missed shipping delay complaints entirely because they appeared in the last 20% of the text, and this limitation is not flexible.

The same models perform differently when called via the Gemini API compared to the Gemini UI, introducing unpredictability in production deployments and making it difficult to validate behavior during development.

Working with domain-specific languages like Terraform requires excessive prompt engineering with Gemini CLI. The model struggles with DSL semantics, necessitating detailed and repetitive prompt tweaking to achieve correct results.

The Gemini API is significantly more verbose and nested compared to competing APIs (Anthropic, OpenAI), making implementation more difficult and time-consuming. The overall design is developer-unfriendly compared to alternatives.

Gemini API's training data cutoff at late 2024 means it cannot answer questions about 2025 technology launches and recent developments. The model returns blank or inaccurate responses for current events.

Developers encounter mysterious failures when working with structured outputs (schemas/grammars) but cannot determine the actual limits causing the failures. The documentation does not clearly explain constraints, making it impossible to debug or optimize queries effectively.

Gemini API returns vague error messages like '500 Internal Server Error' with no context or details about root cause, making debugging extremely time-consuming. One developer spent an entire afternoon debugging straightforward requests due to unhelpful error information.

Token-per-minute (TPM) limits, while less publicized, create additional constraints on large context operations. Developers processing lengthy documents or maintaining extensive conversation histories can hit TPM limits even when RPM and daily request limits are not exceeded.

Developers must implement robust error handling covering multiple HTTP status codes (400, 403, 429, 500) with different retry strategies. Implementing exponential backoff and graceful error catching adds complexity to error handling logic.

Gemini API requires internet connectivity with no local or offline alternative. Applications cannot function without a network connection.

Safety settings in the Gemini API have unexpected behavior and are inadequately documented compared to other vendors, causing confusion during implementation.

Developers encounter frequent `400 INVALID_ARGUMENT` errors when request parameters don't match the API reference, often due to typos or using newer API parameters with older endpoints. The documentation doesn't clearly convey which parameters are required vs. optional for different endpoints.

Google offers Gemini API access through two confusing product names (AI Studio and Vertex AI) with unclear differences for developers. This naming confusion adds unnecessary cognitive load during onboarding and increases time to first API call.