Google Gemini
Listed at https://ai.google.dev
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 2.5 Pro: $1.25-10/M, 2.5 Flash: $0.15-1.25/M, 2.0 Flash: $0.10/M | per million tokens |
| Output | 2.5 Pro: $5-40/M, 2.5 Flash: $0.30-5/M, 2.0 Flash: $0.40/M | per million tokens |
🤖 Supported Models (4)
✨ Pros
- ✓Gemini 2.5 Pro with 1M+ token context window
- ✓Native multimodal (images+audio+video)
- ✓Google Search Grounding for real-time info
- ✓Generous free tier (2.5 Flash sufficient for daily use)
- ✓Competitive Gemma open-source model series
- ✓Vertex AI for enterprise deployment
⚠️ Cons
- ×China requires proxy
- ×API not fully OpenAI-compatible
- ×Large context window but effective utilization questioned
- ×Complex pricing structure (different rates per context length)
- ×Gemini 2.5 Pro severely output-rate limited
🎯 Best For
Ultra-long document processing; multimodal understanding; real-time search augmented apps; generous free tier scenarios
💰 Pricing & Plans
| Model | Context | Input ($/M tokens) | Output ($/M tokens) | Best For |
|---|---|---|---|---|
| Gemini 2.5 Pro (≤200K) | 1M | $1.25 | $10.00 | Long-doc reasoning, complex code |
| Gemini 2.5 Pro (>200K) | 1M | $2.50 | $15.00 | Full-context analysis, premium tier |
| Gemini 2.5 Flash | 1M | $0.15 | $0.60 (no-reasoning) / $3.50 (reasoning) | High-volume, balanced quality/cost |
| Gemini 2.0 Flash | 1M | $0.10 | $0.40 | Cheapest Gemini text, batch workloads |
| Gemini 2.0 Flash Lite | 1M | $0.025 | $0.10 | Ultra-low-cost, simple classification |
| Gemini Embedding | 8K | — | $0.15 (per 1M chars output) | Vector embeddings for RAG |
| Imagen 4 | — | — | $0.04 / image | Standard image generation |
| Veo 3.1 | — | — | $0.10 / second | Short video generation |
| Gemma 3 (open weights) | 128K | self-hosted | self-hosted | On-prem, fine-tunable, no API cost |
🔧 API & Developer Experience
- •API Style: Google Gen AI SDK (Python/Node/Go/Java/REST) and Vertex AI endpoint — NOT OpenAI-compatible by default. OpenAI-compatible wrappers exist via OpenRouter, LiteLLM, or self-hosted proxies.
- •Base URL: https://generativelanguage.googleapis.com/v1beta (Google AI Studio) or https://{region}-aiplatform.googleapis.com/v1 (Vertex AI).
- •SDK Compatibility: Official google-genai SDK unifies Gemini + Imagen + Veo under one import; community SDKs via OpenAI-compatible wrappers.
- •Multimodal Input: Single request accepts text + image (inline or File API) + audio + video (up to 60 min in 1M context) — no separate per-modal endpoints.
- •Streaming: SSE token streaming for text; file/video outputs stream generation progress; audio output synthesis available on Gemini 2.5 Flash.
- •Context Window: 1M tokens on Gemini 2.5 Pro / Flash — largest among commercial text APIs in 2026; Gemma 3 supports 128K.
- •Tool / Function Calling: Native function calling with JSON-schema; code execution sandbox via the code_execution tool; Google Search grounding built in (returns citations).
🔍 Google Search Grounding
Google Search Grounding is the most distinctive capability of the Gemini API: any Gemini 2.x request can opt into real-time web search, and the model returns its answer together with the source URLs it cited. Unlike retrieval-augmented generation systems that require developers to set up a vector database and ingest documents, grounding pulls fresh data from Google's index on every call — useful for fact-heavy assistants, market research bots, news summarization, and 'what changed this week' workflows. Grounded responses also include a 'search entry point' that lets end-users inspect the exact queries the model ran, which simplifies trust-building in customer-facing products. Pricing is per prompt: ground searches on 2.5 Pro cost roughly $35 per 1,000 grounding prompts (separate from token costs), which makes it viable for B2B but warrants budgeting for high-volume consumer apps.
🌐 Regional Availability & Latency
Google AI Studio and Vertex AI are not directly accessible from mainland China — both the generative-language API and the Imagen/Veo endpoints require a stable proxy or VPN. Latency for users inside China is therefore effectively proxy latency plus 100–300ms typical Google API response time, which makes real-time multimodal chat unreliable on mainland networks. Outside China, Google AI Studio from US/EU typically returns first-token in 200–600ms for text and 1–3s for image/video generation starts. The free tier (2.5 Pro 5 RPM, 2.5 Flash 15 RPM, 2.0 Flash 15 RPM, Imagen 4 10/day) is generous enough for prototype validation without a credit card, but production users in China typically route through a third-party aggregator (OpenRouter, LiteLLM-hosted proxies, or an authorized reseller) to avoid operating their own proxy fleet. Developers building for both China and global audiences should expect to maintain two parallel integration paths — direct Google API outside China, aggregator-routed inside.