Gemini 4 Hydrogen: Pricing and API Guide

Live as of November 13, 2026

One sentence: gemini-4-hydrogen is live on TokenRa and callable right now on /v1/chat/completions. The table below has everything you need before writing a single line of integration code.

Reasoning tokens appear in completion_tokens_details.reasoning_tokens. They count toward max_tokens and are billed at the output rate.

Hydrogen, Argon and Carbon: which one is which

Gemini 4 has been discussed under a set of element codenames, which is why searches mix them up. They are not three tiers of one released product — at the time of writing only some of these exist as published models, and Hydrogen is the one that does not. Treat the codename as a label for a slot in the family, not as a spec sheet.

Gemini 4 Hydrogen is now callable. Argon and Carbon are not yet enabled on TokenRa; check the model catalog for updates.

• Model ID is gemini-4-hydrogen. That is the exact string to send in the model field. No version suffix, no other alias.
• Reasoning tokens count toward max_tokens. The model thinks before answering. If you set a tight limit, leave room for both the thinking and the final answer.
• Reasoning is billed as output. Reasoning tokens appear in completion_tokens_details.reasoning_tokens and are charged at the $7.50 / 1M output rate, not a separate reasoning rate.

Quick start

Three steps from zero to a working call. The model ID is gemini-4-hydrogen and the endpoint is the standard OpenAI Chat Completions one, so any OpenAI-compatible client works with a base-URL swap.

curl https://tokenra.io/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"gemini-4-hydrogen","messages":[{"role":"user","content":"Summarise what an API gateway does in two sentences."}]}'

The same adapter also serves gemini-3.8-flash; see Gemini 3.8 Flash API docs for the shared parameter list and limits.

Parameters and behavior

These are the behavior notes that actually matter once your call works. Each one is from a live request made on November 13, 2026, not from a spec sheet.

• Reasoning model. Hydrogen thinks before answering. The reasoning text comes back in reasoning_details[0] with format: "google-gemini-v1". A second reasoning.encrypted entry is Google's opaque thinking payload — you cannot read it, and you should not log or store it.
• Reasoning tokens count toward max_tokens. A call that returns 40 completion_tokens can have 35 of them as reasoning_tokens. If you set a tight max_tokens, the model may run out of budget before it writes the final answer and return an empty content. Leave headroom.
• Billed as output. Reasoning tokens are charged at the $7.50 / 1M output rate, not at a separate or discounted reasoning rate.
• finish_reason: stop (lowercase). The gateway normalizes it to the OpenAI standard, so you do not need a special case for an uppercase STOP.
• No function calling. Requests that include a tools field return 400. Send plain text prompts only.
• Set a long client timeout. Reasoning models can take a while. Set the client timeout to 180 seconds or more and retry once on a 504.

FAQ

Related