Skip to main content
Gemini 4 Hydrogen (gemini-4-hydrogen) is Google’s new model for everyday coding and agent work, and its recommended upgrade from Gemini 3.8 Flash. Since November 14, any default-group LaoZhang API key can call it at $1.5 input and $7.5 output per 1M tokens, the same as Google’s price.
  • Published: November 14, 2026
  • Last checked: November 14, 2026
  • Status: Live in the default group

Key facts

Pricing

Prices from LaoZhang API’s public pricing on November 14, 2026, in US dollars per 1M tokens:
  • There are no context tiers. A prompt over 200K tokens costs the same per token, unlike Gemini 3.1 Pro.
  • Thinking tokens are billed as output. For example, a request with 43 input tokens and 135 output tokens (answer plus thinking) costs about $0.0011.
  • Hydrogen thinks at high unless you say otherwise, so even easy questions produce thinking tokens. Send low when a short answer is enough; see Thinking levels.
  • Cache reads come from Google’s implicit caching: when you resend the same prefix of 4,096 tokens or more, the cached part is billed at the cache price with no extra parameters. Hydrogen does not support explicit caching (cachedContents), so drop that code when you move over from Gemini 3.8 Flash.
  • Per-request token usage and charges are in your usage logs. Current prices are in the model catalog.

Hydrogen vs Gemini 3.8 Flash and Gemini 4 Argon

  • Coming from Gemini 3.8 Flash: Google says Hydrogen is clearly stronger on coding and agent tasks and finishes them faster. Per token it costs twice as much today. Google’s 3.8 Flash launch price runs through December 31, 2026; from January 1, 2027 both models have the same list price.
  • Compared with Gemini 4 Argon: in the benchmark index Google cited at launch, Hydrogen at its default high level scores slightly above Argon, while its per-token price is 25% lower. Once Argon moves to Google’s list price on December 1, Hydrogen costs less than half as much. For the hardest coding, debugging, and science problems, also try Gemini 4 Carbon.
  • The quality claims above are Google’s. Before you switch production traffic, run your own prompts and compare results, latency, and token usage in your usage logs. All three models are called the same way; only model changes.

Request examples

Create an API key under Tokens: choose the default group and set billing to “usage first” (recommended) or “usage”. Then set an environment variable:
  • On the OpenAI-compatible API, the answer is in choices[0].message.content and the thinking token count is in usage.completion_tokens_details.reasoning_tokens.
  • On the native API, the answer is in candidates[0].content.parts[].text, the thinking token count is in usageMetadata.thoughtsTokenCount, and modelVersion reads gemini-4-hydrogen.
  • Code that already calls Gemini 3.8 Flash only needs gemini-3.8-flash replaced with gemini-4-hydrogen in model or the URL. Keep the same API key, and remove explicit caching if you use it.
  • For native image, streaming, and multi-turn requests, see the Gemini protocol. For tools, see function calling.

Thinking levels

  • Without a level, Hydrogen uses Google’s default, high. That is more thinking than Gemini 3.8 Flash, which defaults to medium.
  • xhigh is new with Hydrogen. It usually spends several times the thinking tokens of high, and Google warns it can overthink simple prompts. Save it for hard problems.
  • To use xhigh, or to control the thinking level precisely, call the native API. On the OpenAI-compatible API, stick to the three values in the table.
  • minimal returns an error on the native API. The lowest level is low.
  • Thinking and the visible answer share maxOutputTokens (max_tokens on the OpenAI-compatible API). If the limit is too small, the model can run out before it writes the answer, and the thinking tokens are still billed. To cut cost, lower the thinking level instead of the output limit.
  • Google says Gemini 4 Hydrogen ignores temperature, top_p, and top_k. Sending them does not cause an error, but it does not change the output.

Who this affects

  • Affected: projects using Gemini 3.8 Flash for coding, agents, or batch text work that want a stronger model, and projects that want a cheaper per-token price than Gemini 4 Argon. Any default-group key billed by usage can call it.
  • Not affected: no existing model changed with this launch. gemini-3.8-flash, gemini-4-argon, gemini-4-carbon, and the other Gemini models keep working, and current integrations need no changes.

FAQ

What is the Gemini 4 Hydrogen model ID?

gemini-4-hydrogen. Google published only this GA ID, with no -preview or dated version. Put it in the model field on the OpenAI-compatible API, or in the URL path on the native API.

Google has a free tier. Why call Hydrogen through LaoZhang API?

Google’s free-tier limits apply per project, and Google uses free-tier data to improve its products. On LaoZhang API you use a LaoZhang API key and the base URL https://api.laozhang.ai, with no Google account or AI Studio API key. You pay from a prepaid balance by tokens, at Google’s paid-tier price.

Is Gemini 4 Hydrogen more expensive than Gemini 3.8 Flash?

Per token, twice as much right now: $1.5 input and $7.5 output for Hydrogen versus $0.75 and $3.75 for 3.8 Flash. The 3.8 Flash price is Google’s launch price, which ends December 31, 2026; from January 1, 2027 Google lists both at the same price. What a request actually costs also depends on thinking tokens, and low cuts them sharply.

Why are output tokens much larger than the answer?

Output tokens include thinking tokens. Hydrogen thinks at high by default, even on simple questions. The thinking part is usage.completion_tokens_details.reasoning_tokens (usageMetadata.thoughtsTokenCount on the native API). Send low when a short answer is all you need.

Does it accept images, PDFs, audio, or video?

Image input works: use image_url (URL or Base64) on the OpenAI-compatible API, or inlineData / fileData on the native API. Google also lists video, audio, and PDF input; see image understanding and the Gemini protocol. The model outputs text only, not images or speech.

What if long requests time out?

At the default high level a request can take tens of seconds, and xhigh takes longer. Set your client timeout to about 600 seconds and stream the response:
  • OpenAI-compatible API: add "stream": true to the request.
  • Native API: call streamGenerateContent?alt=sse instead.