模型與開發者平台
Gemini 3.7 Flash Targets Coding Agents; LiteLLM Warns That the `minimal` Reasoning Level Returns HTTP 400
Google is rolling out Gemini 3.7 Flash across its API, Antigravity, Android Studio, and enterprise agent platforms, offering a limited-time 50% discount to capture high-volume agent workloads. Early integrations show that multimodal input, tool calling, and cross-API formats are already supported, but reasoning parameters and cost data still pose migration pitfalls.
Google launched Gemini 3.7 Flash on August 13, positioning it not as a smaller chat model but as a high-throughput “workhorse model” for coding, multi-step agents, and complex knowledge work. It is now available through the Gemini API, Google AI Studio, Antigravity, Android Studio, and enterprise agent platforms. Gemini Spark has also switched to the model for tool-based workflows involving Gmail, Calendar, Docs, and other services.
For engineering teams, the more concrete changes are in the interfaces rather than the product name. LiteLLM has confirmed that the model can be called through Google AI Studio’s `gemini/` route or Vertex AI’s `vertex_ai/` route. It supports `/v1/chat/completions`, `/v1/responses`, the Anthropic-compatible `/v1/messages`, and the native `generateContent` API. Streaming, function calling, thought signatures, multi-turn conversations, and text, image, audio, and video inputs are all supported, allowing existing agents to switch models through the same gateway.
Reasoning parameters still cannot be carried over without adjustment. LiteLLM maps the OpenAI-format `reasoning_effort` parameter to Gemini’s `thinkingLevel`, but the initial Gemini 3.7 Flash release does not accept `minimal`; requests using it return HTTP 400. Before deployment, teams should filter parameters according to the model capability table instead of allowing retry logic to repeatedly submit the same invalid request. Older LiteLLM versions can run inference directly, but cost tracking requires the model cost map to be reloaded. Deployments using a locally pinned cost table need at least the data included with `v1.98.0-dev.2`.
Through the end of 2026, promotional pricing is $0.75 per million input tokens and $3.75 per million output tokens, with both rates scheduled to double afterward. This makes the model particularly suitable for large numbers of subagents and tool loops, although Google’s claimed gains in coding and instruction following still need to be validated against real repositories, long conversations, and recovery from tool failures. Teams should next watch for `minimal` support, a stable model alias, an official cost table, and whether longer outputs at higher reasoning levels offset the model’s per-token price advantage.