Gemini 3 API now available for developers: 1M token context, Generative UI, 76.2% SWE-bench coding, Google Antigravity IDE. $2/M tokens. Build full-stack app...
Gemini 3 Is Now an API You Can Actually Call
Google has opened Gemini 3 to developers through its API, which shifts it from something you read about to something you can wire into your own software today. The headline capability is a 1M token context window, meaning you can hand the model an entire codebase, a long document set, or a lengthy transcript in a single request rather than chopping it into fragments and stitching the answers back together.
A context window this large changes how you design prompts. Instead of building elaborate retrieval systems to feed the model small, carefully chosen snippets, you can often just include the whole source and let the model find what matters. That said, more context is not free — larger inputs cost more tokens and can slow responses — so it pays to include what's genuinely relevant rather than everything you have lying around.
Coding, Generative UI, and the Antigravity IDE
Gemini 3 posts a 76.2% score on SWE-bench, a benchmark built from real software engineering tasks pulled from open-source projects. A number like that signals the model is aimed at practical coding work — reading an existing repository, understanding an issue, and proposing a concrete fix — rather than only toy snippets. The summary also points to being able to build a full-stack app, which lines up with that coding focus.
Two features extend this beyond raw text generation. Generative UI means the model can produce interface elements, not just prose or code strings, which is useful when you want a working front end rather than a description of one. Google Antigravity is the IDE built around these capabilities, giving you an editor where the model is a first-class participant in the loop instead of a separate chat window you copy and paste between.
What $2 Per Million Tokens Means for Your Budget
Pricing is listed at $2 per million tokens. Token-based pricing charges you for the text going in and coming out, so your real cost depends on how much context you send and how much you ask the model to generate. The 1M context window and this price interact directly: filling the window on every call is powerful but adds up, so treat large-context requests as a deliberate choice.
A few practical habits keep spending predictable:
- Send only the context a task needs — trim boilerplate and irrelevant files before the request.
- Cache or reuse prompts where the same context is queried repeatedly.
- Estimate token counts for your typical request before scaling up, so a per-call cost turns into a per-day cost you can plan around.
How to Start Building
The fastest path is to pick one concrete task you already understand well — a small feature, a bug fix, a document summarizer — and implement it against the API end to end. Working from a task you know lets you judge the output quality honestly instead of guessing. Use the large context window to give the model real material from your project, and lean on the coding strength and Generative UI when the goal is a running app rather than a one-off answer.
If you prefer an integrated setup, start inside Antigravity so the model sits next to your code from the beginning. Either way, keep your first project small enough to finish, measure the token cost of a typical run, and expand only once the results and the economics both hold up.