AI Tools12 min read

Gemini 4 Hydrogen vs Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna: Price, Speed and Thinking Levels

Gemini 4 Hydrogen costs $1.50/$7.50 per million tokens. Cost per task against intelligence at every thinking level, compared with Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna.

By Moe Lueker·Nov 12, 2026

Google released Gemini 4 Hydrogen today, November 12, 2026. It costs $1.50 per million input tokens and $7.50 per million output tokens, and it is the new default fast model in the Gemini app and Google Search.

That price is twice what Gemini 3.8 Flash costs right now and fifteen times Claude Haiku 5.5 and GPT-6 Luna. So the useful question is not the sticker price. It is how much intelligence you get per dollar of real work, and which thinking level gets you there. I pulled the numbers for every thinking level of all four models from Artificial Analysis, plus Google's announcement, model card and pricing page.

If you only have a minute#

What does it cost? $1.50 in and $7.50 out per million tokens, flat across the full 1 million token context window. Cached input is $0.075. Batch and Flex are half price.

Is it smarter than the others? By a lot. It scores 54 on the Artificial Analysis Intelligence Index at xhigh. The best the others reach is 36 (Haiku 5.5 on max), 31 (3.8 Flash on high) and 28 (Luna on max).

Is it more expensive per task? Than Haiku and Luna, yes. Than Gemini 3.8 Flash, no: Hydrogen on high costs $1.87 per Index task, while 3.8 Flash on high costs $7.77 and scores 20 points lower.

Which thinking level? Medium is the value pick: 41 on the Index for $0.80 a task, already above every setting of the other three. Use high (the default) for hard coding and agent work. Xhigh adds 3 points for almost three times the thinking tokens.

Is it fast? Independent speed tests are not out yet. Google says it finishes tasks 2.5 times faster than 3.8 Flash.

When should you use something else? For high-volume, simple jobs like sorting, extraction and short replies, Haiku 5.5 and Luna do the work for a few cents a task.

Gemini 4 Hydrogen price vs Gemini 3.8 Flash, Haiku 5.5 and Luna#

All prices are per million tokens on the standard paid tier, checked November 12, 2026.

Three things the table doesn't show on its own:

• Gemini 3.8 Flash is about to cost the same. Google's pricing page lists Flash at $0.75 and $3.75 only until December 31. From January 1, 2027, it is $1.50 and $7.50, exactly Hydrogen's price. After that there is no reason to stay on Flash.
• Haiku 5.5 has a price cliff. Its $0.10 and $0.50 rate covers prompts up to 100,000 tokens. Above that, the whole request costs five times more. I covered that in my Claude Haiku 5.5 price and benchmarks post.
• Hydrogen has no long-context surcharge. The $1.50 rate holds across the full window. It only supports implicit caching, though, so you can't pin a cache yourself.

Cost per task vs intelligence at every thinking level#

Token price times tokens used is what you pay. A model that thinks for longer at a lower price can cost more per task than a pricier model that gets to the point. Artificial Analysis runs every thinking level of every model through the same set of nine evaluations and reports the Intelligence Index and the average cost of one task. Here is all of it on one chart:

What stands out:

• Hydrogen is in a different tier. Even its medium setting, at 41, beats the best score of every other model here.
• Hydrogen replaces 3.8 Flash outright. Hydrogen on medium scores 10 points higher than Flash on high for about a tenth of the cost per task.
• Under 50 cents a task, Haiku 5.5 is the one to beat. Haiku on high scores 30 for $0.16. Hydrogen on low scores 27 for $0.29. If your budget is that tight, Haiku gives you more.
• GPT-6 Luna is the cheapest floor. One to fifteen cents a task, but it tops out at 28.

These are Index v5.0.1 numbers. My earlier Opus 5.5 vs GPT-6 Sol and Luna comparison used an older version of the Index, so the scores there are not on the same scale.

How the curves change across thinking levels#

Every model here lets you turn thinking up or down. Each step up spends more output tokens, and output tokens are where the money goes. The curves look very different from model to model:

The steps for Gemini 4 Hydrogen:

The first two steps pay off. The last one mostly buys thinking. Google's own developer docs warn that xhigh "can cause overthinking on simpler prompts."

The other models flatten out the same way, just lower down. Haiku 5.5 gains 1 point from xhigh to max while its wait for the first answer token jumps from about 52 seconds to 183. Luna's best jump is xhigh to max, 5 points for 50% more cost. Gemini 3.8 Flash barely moves: medium and high are 30 and 31.

Speed: what we know on launch day#

Artificial Analysis had not measured Hydrogen's speed when I wrote this, so there is no independent tokens-per-second number yet. Here is what is measured for the others, plus Google's claim:

Google's speed claims, from the announcement:

• Finishes tasks 2.5 times faster than Gemini 3.8 Flash at the default high setting.
• Partners measured 2 to 2.8 times faster than 3.8 Flash, with 12% to 23% fewer tokens.
• The model card lists "occasional slowness or timeout issues" as a known limitation.

Those are task-level claims, not streaming speed. Fewer thinking tokens alone would make a task finish sooner. I will add the independent numbers when Artificial Analysis publishes them.

If raw streaming speed is what you need today, Claude Haiku 5.5 is the clear winner at 216 to 256 tokens per second, more than twice Flash and Luna. Just keep it off max if you are waiting on it: the time to the first answer token there is about three minutes.

Benchmarks from Google's model card#

Google did not compare Hydrogen with Haiku 5.5 or Luna. Its model card puts it against Claude Sonnet 5.5 and GPT-6.1 Terra, both bigger, pricier models, and does not say which thinking level each score used.

Against its own predecessor, Google reports 57% on Terminal-Bench 4.1 for Hydrogen and 18.2% for Gemini 3.8 Flash. That is the biggest single jump in the launch, and it lines up with the Intelligence Index gap above. Sonnet 5.5 still wins on terminal work and computer use, which is worth knowing if those are your main jobs.

These are Google's numbers. The Artificial Analysis results above are the independent ones, run the same way for every model.

Which model to use for what#

Use Gemini 4 Hydrogen for:

• Coding and debugging across a repository, on high
• Agent tasks and business workflows, where it leads AutomationBench-AA and Vending-Bench 2
• Long documents, PDFs, video and audio in one prompt, up to 1 million tokens
• Anything you used to run on Gemini 3.8 Flash. Hydrogen on medium is smarter and cheaper per task
• Everyday writing and research in the Gemini app, where it is now the default

Use Claude Haiku 5.5 for:

• Cheap, fast subagents and bulk work under 50 cents a task
• Anything where streaming speed matters
• Short prompts. Keep them under 100,000 tokens to avoid the 5x price tier

Use GPT-6 Luna for:

• The lowest possible cost per task: triage, tagging, extraction, routing
• High-volume pipelines where a score in the 20s is enough

Stop using Gemini 3.8 Flash for new work. Its launch price ends on December 31, and Hydrogen beats it on every independent number measured so far.

How to get good results from Gemini 4 Hydrogen#

1. Set the thinking level yourself. It is thinking_level in the API: low, medium, high or xhigh. There is no minimal, and sending it returns an error.
2. Start on medium and move up only when it falls short. Most tasks don't need the default high.
3. Leave temperature alone. Google's docs say temperature, top_p and top_k are ignored on this model.
4. Use Batch or Flex for anything that can wait. Both cut the price in half.
5. Write a clear brief. Say what the task is, what context matters, what to skip and what a good answer looks like. A vague prompt makes any thinking model spend tokens guessing. I speak my briefs with Rambleproof, the Mac dictation app I built, and the clean text lands in Gemini, Claude or wherever my cursor is.

Spend fewer tokens with a better prompt#

The cheapest way to cut a thinking model's bill is not a lower thinking level. It is a prompt that answers the questions the model would otherwise think through on its own. Most people type short prompts because typing a long one is tedious. Talking isn't.

Rambleproof lets you press Right Option, say what you want, press it again, and get clean text in Gemini, ChatGPT, Claude, your editor or the other apps you type in. It cleans up the ums and restarts. It runs on your Mac and has a free plan.

Quick answers#

The short version#

Gemini 4 Hydrogen came out on November 12, 2026, at $1.50 in and $7.50 out per million tokens. It is the smartest model in this price class by a wide margin: 54 on the Artificial Analysis Intelligence Index at xhigh, against 36 for Claude Haiku 5.5, 31 for Gemini 3.8 Flash and 28 for GPT-6 Luna.

Run it on medium for everyday work and high for hard work. It replaces Gemini 3.8 Flash completely. Keep Haiku 5.5 and Luna for cheap, high-volume jobs where a few cents a task matters more than the last bit of intelligence.

If you are weighing the bigger models too, my Claude Fable 5.2 vs Opus 5.5 post covers the top end with the same cost-per-task approach.

Published November 12, 2026, the day of release. Sources: Google's Gemini 4 Hydrogen announcement, the developer post, the model card, the Hydrogen model docs, Google AI Studio pricing, Claude pricing and the Artificial Analysis pages for Gemini 4 Hydrogen, Gemini 3.8 Flash, Claude Haiku 5.5 and GPT-6 Luna, with each effort level on its own page. The full dataset is at /data/gemini-4-hydrogen-vs-flash-haiku-luna.csv. Not sponsored by Google. I will update this post when independent speed tests land.