Gemini 4 Hydrogen exposes four thinking levels, and the request surface is narrower than the list suggests. thinking_level accepts low, medium, high and xhigh; the default is high. thinking_budget is not an alternative: on any Gemini 4 model it returns 400: INVALID_ARGUMENT. MINIMAL is rejected, custom sampling values are ignored, and AIHubMix does not document how to set a thinking level for any Gemini 4 model.
The spread is wide enough to plan around. On Artificial Analysis's live measurements, the same model scores 27 at low, 41 at medium, 51 at high and 54 at xhigh — 27 index points inside one model ID. high is the knee: the last step where the quality gain is large relative to the added cost. xhigh is for work where high fails.
The request surface
A working request needs the AIHubMix base URL, a key in AIHUBMIX_API_KEY and the model ID gemini-4-hydrogen. The documented route is POST https://aihubmix.com/v1/chat/completions.
The parameter is thinking_level in snake_case. Google's Python sample sets it inside thinking_config; the Node sample nests it under thinkingConfig in camelCase. The model page lists the values as LOW, MEDIUM, HIGH and XHIGH; the release notes write them lowercase. Use the SDK's enum or the documented string rather than hand-typing a case variant.
MINIMAL is not part of the set. The Gemini 4 thinking documentation says MINIMAL is "not supported on Gemini 4 models," and there is no documented off switch for thinking on this model.
| Surface | Documented behaviour |
|---|---|
| thinking_level | low, medium, high, xhigh; default high |
| thinking_budget | 400: INVALID_ARGUMENT on any Gemini 4 model |
| thinking_level with thinking_budget | error |
| thinking_level=MINIMAL | not supported; Hydrogen rejects it |
| temperature, top_p, top_k | ignored; fixed at 1.0, 0.95 and 64 |
| max_output_tokens | caps thinking and output together |
The shape AIHubMix documents leaves the level out:
import os
from openai import OpenAI
client = OpenAI(
base_url="https://aihubmix.com/v1",
api_key=os.environ["AIHUBMIX_API_KEY"],
)
response = client.chat.completions.create(
model="gemini-4-hydrogen",
messages=[{"role": "user", "content": "Summarise this stack trace and name the likely fix."}],
max_tokens=16384,
)
print(response.choices[0].message.content)
Whether thinking_level passes through the gateway is not documented. Verify it with a small request before it goes into a production client.
What changed in the family
Google's release notes put it plainly: "The model defaults to high and introduces xhigh." The same release notes say the model "rejects minimal with an error."
That is two changes against Gemini 3.8 Flash, which defaults to medium and stops at high.
Gemini 4 Carbon and Argon did not move. Both support low, medium and high, with medium as the default. Hydrogen is the only Gemini 4 model with a setting above high.
| Model | Levels | Default |
|---|---|---|
| Gemini 4 Hydrogen | low, medium, high, xhigh | high |
| Gemini 4 Carbon | low, medium, high | medium |
| Gemini 4 Argon | low, medium, high | medium |
What the four levels score
Artificial Analysis measured all four levels. The run is Intelligence Index v5.0.1, 78 ranked models, accessed at 2026-11-13T02:40Z, at standard list prices. These are AA task costs, not AIHubMix bills.
| Level | AA rank | AA Index | $/task | Min/task | First answer token |
|---|---|---|---|---|---|
| low | 60 | 27 | 0.29 | 1.3 | 5.71 s |
| medium | 39 | 41 | 0.80 | 3.7 | 11.80 s |
| high | 21 | 51 | 1.87 | 8.3 | 18.89 s |
| xhigh | 16 | 54 | 3.16 | 13.5 | 37.77 s |
The step sizes are the useful part:
-
lowtomedium: +14 index points, +$0.51 per task, +2.4 minutes, +6.1 seconds to the first token. -
mediumtohigh: +10 index points, +$1.07 per task, +4.6 minutes, +7.1 seconds to the first token. -
hightoxhigh: +3 index points, +$1.29 per task, +5.2 minutes, +18.9 seconds to the first token.
high is the knee. It is the last step where the quality gain is large relative to the added bill; xhigh is the step where the first-token wait nearly doubles for a small index gain. That does not make xhigh a bad setting. It makes it a setting for tasks where high fails, which is how Google describes it.
For context in the same snapshot: Carbon high scores 60 at $5.48 per task and 28.1 minutes, Carbon medium 58 at $4.68, Carbon low 54 at $5.22, and Argon high 50 at $4.97 and 30.6 minutes. Hydrogen xhigh matches Carbon low's 54 at 39.5% lower cost per task ($3.16 against $5.22), though its first token takes 37.77 seconds against 13.61. Hydrogen high beats Argon high by one index point at 62.4% lower cost per task.
Where requests fail
A thinking_budget fails the request. Any Gemini 4 model returns 400: INVALID_ARGUMENT, and setting both thinking_level and thinking_budget also returns an error. A client library that sends a budget by default fails the request before the model runs. These are Google's documented responses; no live test was run against AIHubMix for this post.
MINIMAL does not fall back. Hydrogen rejects it with an error, so a config that treats MINIMAL as "off" breaks rather than degrades.
Sampling values are ignored. temperature, top_p and top_k are accepted and discarded; the fixed defaults are temperature 1.0, topP 0.95 and topK 64. A prompt tuned against a low temperature does not behave the same way after migration.
Long output hits a hop limit. A synchronous request generates up to 65,536 tokens per hop, streaming or not. Longer output needs background execution or continuation tokens, and max_output_tokens caps combined thinking and output. If background=true is set without a max_output_tokens, the server stops after 65,536 tokens. Google's suggested long-decode budgets are 131,072 or 262,144, up to 1,048,576 minus the input.
xhigh can overthink. Google says "Using xhigh increases latency and token use and can cause overthinking on simpler prompts," and the measured figures show the shape of that: $3.16 per task, 13.5 minutes and a 37.77-second first-token wait.
How thinking becomes the bill
On Google's rate card, Hydrogen costs $1.50 per million input tokens, $7.50 per million output tokens and $0.075 per million cached input tokens. Output includes thinking tokens, so the level changes the bill through the tokens it writes, not through a surcharge on the level name.
Three rules matter at migration.
The default changed. Gemini 3.8 Flash defaulted to medium; Hydrogen defaults to high. A workload that never set a level spends more output tokens per turn after migration, at the same per-token price. That is a budget change disguised as a model swap.
Caching is implicit only. Hydrogen supports implicit caching with a 4,096-token minimum and a 95% discount on cache hits; explicit caching is not supported. Google's enterprise page says "cache writes are billed at 1.25x the standard input token price by default." That is Google's billing wording; AIHubMix's cache billing is not documented for this model.
Verbosity is part of the bill. Artificial Analysis describes Hydrogen as "very verbose": the xhigh run generated 304 million output tokens to complete the index, rank #76 of 77, against a median of 72.8 million. Output is billed at $7.50 per million. A higher level does not just think longer; it produces more billable tokens.
What AIHubMix documents
AIHubMix's documented route for Hydrogen is POST https://aihubmix.com/v1/chat/completions with model="gemini-4-hydrogen". The model page lists text, image, audio, video and PDF input with text output; llms.txt still carries an older text-only line, so treat the page as the newer source. The capability flags for tool use, structured output and reasoning are "not published." The model is not in the model index generated on 2026-11-11, and no protocol schema is published.
The gap is specific: AIHubMix does not document how to set a thinking level for any Gemini 4 model. Its Gemini guide covers reasoning_effort for Gemini 2.5-era models, not thinking_level or xhigh for Gemini 4. Do not assume that either parameter passes through the gateway, and do not assume the default observed on Google's API is the default you get on AIHubMix.
If level control is a hard requirement, call Google's API directly or verify the gateway's behaviour with a small request before routing production traffic. The AIHubMix compare page is a cost-side check while you decide. No live test was run for this post.
Starting points by workload
| Workload | Start at and why |
|---|---|
| Fast chat, routing, classification | medium, test low; low scores 27 |
| Interactive coding, UI prototyping | high; AA's medium is 41 |
| Agent execution, tool loops | high; xhigh only if stuck |
| Hard debugging, architecture, planning | xhigh or Carbon; 3 points for 69% more |
| High-volume batch after evaluation | low or medium; cheapest, but 27/41 index |
| Strict latency budget under 15 seconds | medium or low; high waits 18.89 s |
These are starting points, not routing rules. The right level is the lowest one that passes your own evaluation, and Google's advice is to run that evaluation at all four levels with a high max_output_tokens.
Keep reading: the Gemini 4 Hydrogen series
- If the question is whether Hydrogen is worth the move from Gemini 3.8 Flash at all, the first post covers the launch, the price change and the benchmark trade-offs: Gemini 4 Hydrogen vs Gemini 3.8 Flash: Is the Upgrade Worth It?
- If the question is Hydrogen against the rest of the family, the third post puts Carbon, Hydrogen and Argon side by side with prices, latency and a workload table: Gemini 4 Carbon vs Hydrogen vs Argon: Which Model for Which Job?
Sources
- Google AI Studio, Gemini 4 Hydrogen: what's new
- Google, Gemini 4 thinking
- Google, Gemini 4 Hydrogen model page
- Google AI Studio, Gemini API pricing
- Artificial Analysis, Gemini 4 Hydrogen — live snapshot 2026-11-13T02:40Z, Intelligence Index v5.0.1
Top comments (0)