
Gemini 4 Hydrogen: What xhigh Buys for 6x the Bill
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- openaiNEWOpenAI: GPT-6.2 Astra2026-11-17$8.00 / $40.00 per 1M tokens · 76 tok/s
- openaiNEWOpenAI: GPT-6.2 Sol2026-11-17$2.00 / $10.00 per 1M tokens · 151 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 240 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 178 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 94 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 212 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 131 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 82 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 105 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- openaiOpenAI: GPT-6.1 Terra2026-10-26$1.00 / $5.00 per 1M tokens · 36 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 39 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 222 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 104 tok/s
Two runs of the same short prompt through Gemini 4 Hydrogen, one setting apart. At high the model spent 5,311 output tokens and billed 3.99 cents. At xhigh it spent 31,477 and billed 23.61 cents. The same comparison appears in a Gemini 4 Neon run on the same day, where the top rung cost 39.83 cents against 14.55 at high. Simon Willison published both pairs on 20 November 2026, four days after Google shipped Gemini 4 Hydrogen on 12 November, and the numbers are the first public side-by-side at the very top of a Gemini 4 effort ladder that most people have not yet touched.
The interesting part is not that the deeper setting costs more. It is that the per-token rate never moved. Hydrogen charges $1.50 per million input tokens and $7.50 per million output tokens at every rung, including xhigh and including thinking tokens. Multiply the token counts by that rate and you get the bills back to the cent. So xhigh is not a premium tier you pay extra for. It is a decision to let the model write about six times as much, and the invoice follows automatically.
The two bills, and why the rate held
The arithmetic is worth doing by hand once, because it settles what you are actually buying. Hydrogen's output rate is $7.50 per million tokens. At 5,311 tokens that is 3.98 cents; the measured bill was 3.99. At 31,477 tokens it is 23.61 cents, which is the measured bill exactly. The input side is noise at this scale: a prompt of a few dozen tokens costs a fraction of a hundredth of a cent at $1.50 per million.
• Output tokens — 5,311 at high against 31,477 at xhigh, a factor of 5.9
• Cost in cents — 3.99 against 23.61, the same factor of 5.9
• Output rate — $7.50 per million tokens at both settings, unchanged
• Input rate — $1.50 per million tokens at both settings, unchanged
• Context — a single rate to 1M tokens, with no long-context surcharge listed
That has a practical consequence for anyone budgeting this model. Vendor price sheets are organised by token rate, and the rate is the one variable that does not move here. Planning for xhigh from a per-token table will underestimate the bill by roughly six times on a task where the model chooses to think longer. The only figure that predicts the cost is output tokens per answer, and the only way to get that is to measure it on your own prompts. Google's own documentation concedes the direction without giving a magnitude: it warns that xhigh "increases latency and token use."

What the extra 26,000 tokens produced
On the pelican-on-a-bicycle test the answer is: a visibly better drawing. Willison's summary is that the high output was a flat pelican with no basket and no background, and the xhigh output arrived as a teal bike on a beach with a blue fish in the basket — competitive with what Gemini 4 Neon produces by default at roughly the same money. Roughly 26,000 extra thinking tokens went into composition, scene and colour rather than into a correct answer.
Treat that as a single observation, not a benchmark. It is one prompt, one sample per setting, and a task with no single correct answer, which means no accuracy score can be attached to it. What it does establish is that the top rung can change the character of an output on open-ended generative work, and that the change is not subtle when it happens. Whether the same six-fold spend buys anything on your workload is a separate question, and the next two sections are about the cases where the answer is measured rather than eyeballed.
The independent ladder says the same
Artificial Analysis scores Hydrogen at each of its four settings, and the shape of the curve matches the pelican bill. Reading the model's page on 22 November 2026, the Intelligence Index runs 26.9 at low, 41.2 at medium, 50.8 at high and 53.9 at xhigh. Cost per Index task runs $0.29, $0.80, $1.87 and $3.16 over the same four rungs, and time per task runs 1.5, 4.0, 9.0 and 15.8 minutes.
Turn those into marginal cost — the price of the next point — and the argument against defaulting to the top rung becomes arithmetic rather than taste:
• low to medium — +14.3 Index points for +51 cents, about 3.5 cents per point
• medium to high — +9.6 points for +$1.08, about 11.3 cents per point
• high to xhigh — +3.1 points for +$1.29, about 41.1 cents per point
• Total, low to xhigh — +27 points for +$2.87, about 10.6 cents per point on average
The last step is the expensive one by a factor of nearly four against the step before it, and it also costs 76 percent more wall-clock time per task. The first step from low to medium remains the best purchase in the ladder by a wide margin, which is worth remembering given that high is the default. These are third-party measurements from Artificial Analysis, not vendor figures, and the per-task costs bundle input, output and multi-turn agent behaviour — which is why they do not scale the way a single prompt's bill does.

Where xhigh buys nothing at all
Google's documentation is blunt about the upper rung in a way that launch material rarely is. It says to reserve xhigh "for hard, multi-step coding and agentic problems where high falls short," and warns that on simpler prompts it "can cause overthinking or exhaust max_output_tokens." That second failure mode is the one that bites silently: the token limit covers thinking as well as visible output, so a prompt that would have answered comfortably at high can burn the entire budget reasoning and come back marked incomplete, truncated, or empty. It reads as a model failure. It is a configuration failure.

An independent 56-request sweep published in November 2026 tested the question that Google's warning implies. Across three easy-to-medium prompts and two runs per level, it found xhigh used about 3.1 times the thinking tokens of high overall and about five times as many on the coding prompt, with every answer still correct at every level. The extra spend bought tokens and latency, not accuracy. The sample is small and the author says so; the direction is consistent with both the vendor warning and Artificial Analysis's curve, where the final rung adds three points rather than fifteen.
The split is therefore task class, not model capability. Extraction, classification, formatting and short factual answers have a ceiling close to their first correct answer, and paying six times for it is pure waste. Open-ended generation has no such ceiling, because the output is the artifact and there is always more detail available to add.
The crossover that actually matters
Here is the number that should change how you use the model. Hydrogen at xhigh cost 23.61 cents for one short drawing prompt. Gemini 4 Neon at its default high cost 14.55 cents for the same prompt on the same day, and produced an output Willison judged at least as good. Stepping Hydrogen to its top rung is 62 percent more expensive than leaving it at the default and stepping up a model tier instead. If your reason for reaching for xhigh is "the answer isn't good enough," a bigger model at its default setting is often the cheaper move, and it is the one most teams do not check.
That ordering does not survive a change of workload, which is the part that gets lost when a short-prompt test is quoted as a general ranking. On Artificial Analysis's long agentic basket, Neon at high costs $3.74 per task against Hydrogen at xhigh at $3.16, so the cheaper model on a forty-token drawing is the dearer one on a long multi-turn task. Neither number is wrong. A short prompt and a 15-minute agent run exercise different parts of the ladder, and a rung swap that saves money on one can lose it on the other.
One dated wrinkle to fold in: Google's pricing page lists Carbon and Argon rates doubling on 1 December 2026, from $2 input and $10 output up to $4 and $20. Hydrogen shows no such change. Every comparison above that involves the older models therefore has nine days left before the gap widens in Hydrogen's favour.
Setting it, and testing it in an afternoon
The level is one field. On the native generateContent API it is thinkingConfig.thinkingLevel; on the Interactions API it is generation_config.thinking_level; on an OpenAI-compatible route it arrives as reasoning_effort. The value is the lowercase string — low, medium, high or xhigh — and high is the default for Hydrogen. Thinking cannot be switched off entirely on this model, and the older thinkingBudget parameter is deprecated in favour of the level.
Measuring the cost takes one extra read. Thinking tokens appear in usageMetadata.thoughtsTokenCount on the native API, usage.total_thought_tokens on Interactions, and reasoning_tokens on the compatible route — so you can separate what the model thought from what it said, and price each. A workable afternoon: take twenty real prompts from your own logs, run each at high and xhigh with a generous output limit so nothing truncates, and record tokens and whether the answer passed. Compare total cost against pass rate. If xhigh does not lift the pass rate enough to justify a bill roughly six times larger on the prompts where it engages, you have your answer for that workload, and it will be a better answer than any published grid.
Running two settings side by side is easier when both sit behind one key, which is the part OrcaRouter covers. We serve Gemini 4 Argon and Gemini 4 Carbon at provider list price with 0% markup, so the rates in this article are the rates you pay on our side for those two. Gemini 4 Hydrogen is not in our catalogue today, so every Hydrogen figure here is Google's list price rather than ours — worth saying plainly, because the failure mode in this post is exactly the one where a rate sheet hides a six-fold multiplier. Where a model is routed through us, the effort level is a per-route field, so a canary route at xhigh alongside a production route at high is a config change rather than a code path.
What to do with this on Monday
The rule that survives all the numbers is narrower than "use the highest setting for hard problems." It is: descend the ladder until quality breaks, then take one step back up, and check whether a different model at its default is cheaper than the extra rung before you commit. Hydrogen's own ladder puts that step at roughly forty cents a point on the Index and six times the tokens on a drawing, while the step from low to medium costs three and a half cents a point. Most of the value in this model is in the bottom half of the ladder, and the default sits above it.
The open question is what the top rung does on agentic work rather than on a pelican — whether the 3.1 Index points Artificial Analysis measures at xhigh show up as fewer failed tool calls or as more turns spent thinking about them. Nobody has published that breakdown yet, and it is the one that would tell you whether to leave Hydrogen at high in production.
