Gemini最新模型弃用采样参数

Google发布了Gemini 3.6 Flash和Gemini 3.5 Flash-Lite两款新模型，现已正式可用 1。

API变更

这两款模型已弃用temperature、top_p和top_k采样参数 1。API将忽略这些参数，并在未来版本中对包含这些参数的请求返回HTTP 400错误 1。此外，模型不再支持预填充模型回合，以模型角色回合结尾的API请求将返回400错误 1。

性能与规格

Gemini 3.6 Flash的输出token成本已降至$7.50/1M，相较Gemini 3.5 Flash的$9.00/1M有所下降，输入token成本保持在$1.50/1M 1。Gemini 3.5 Flash-Lite在HLE推理任务上得分为18.0%（相比前版本的11.0%），在CharXIV多模态基准测试中得分74.5%（相比前版本的63.7%） 1。

两款模型均支持1M token上下文窗口和64k最大输出token 1。此外，Gemini 3.6 Flash已成为Gemini Managed Agents中Antigravity agent的新默认模型 1。

---

Google has released two new Gemini model versions that introduce significant changes to API behavior. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are now generally available 1, with the former reducing output token costs to $7.50 per million tokens compared to $9.00 per million for Gemini 3.5 Flash, while input token pricing remains at $1.50 per million 1.

The new models discontinue support for three sampling parameters: temperature, top_p, and top_k 1. These parameters have been deprecated, and the API will ignore them in current requests while returning HTTP 400 errors in future model versions 1.

Both models support a 1 million token context window and can generate up to 64,000 output tokens 1. Additionally, the API no longer permits pre-filled model turns; requests ending with a model role turn will return a 400 error 1.

Gemini 3.5 Flash-Lite shows improved performance on certain benchmarks, scoring 18.0% on HLE inference tasks compared to 11.0% previously, and achieving 74.5% on the CharXIV multimodal benchmark versus 63.7% 1.

The Gemini 3.6 Flash model has been designated as the new default for the Antigravity agent within Gemini Managed Agents 1.

---