Google Gemini API

Google's Gemini API (ai.google.dev) is Google's developer-facing gateway to the Gemini model family, spanning the high-reasoning Pro tier, the cost-efficient Flash and Flash-Lite tiers, and image models. It is offered in two surfaces: the standalone Gemini Developer API via Google AI Studio (the focus here, aimed at individual developers and fast prototyping with a generous free tier) and the enterprise-grade Vertex AI version on Google Cloud (with VPC, data residency, and SLAs). Its signature differentiator is native multimodality (text, image, audio, video, PDF in a single request) paired with a very large context window, up to 1 million tokens on 2.5 Pro/Flash, which has made it a default for whole-codebase analysis, long-document RAG, and video understanding. The official Google GenAI SDK (google-genai) reached GA in May 2025 across Python, JS/TS, Go, and Java, and the API is also exposed via an OpenAI-compatibility layer.

On the cost/performance frontier, Gemini is among the most aggressively priced frontier offerings: 2.5 Flash-Lite runs $0.10/$0.40 per 1M input/output tokens and 2.5 Flash $0.30/$2.50, while 2.5 Pro lists at $1.25/$10.00 (rising to $2.50/$15.00 above a 200K-token prompt). Batch mode halves those, and context caching plus a free tier lower the barrier further. Third-party measurement (Artificial Analysis) puts 2.5 Pro at roughly 132 output tokens/sec, above its category median, but with a high ~21s time-to-first-token typical of heavy reasoning models, and an Intelligence Index score that sits mid-pack rather than top-of-class. Developers consistently praise the 1M context for large-codebase and long-document work; Simon Willison's widely-cited example had it correctly identifying 18 files to change across an entire codebase.

The clearest weakness is reliability. Google's own developer forum has running threads (notably a "2026 Stability Crisis" thread) complaining about elevated 503/overload errors, "infinite thinking" loops, and instability on preview models, with peak 503 rates reported near 45% on some preview image endpoints. Quotas and rate limits on the free and lower paid tiers are also a frequent friction point, and the split between AI Studio and Vertex (different SDKs historically, different quotas) adds confusion. Net: excellent price-to-context ratio and best-in-class multimodality, but buyers needing predictable production uptime tend to route through Vertex AI with its formal SLA rather than the bare Developer API.