Gemini 4 Hydrogen explained: why its price per token says little, and what its cost per task shows next to Carbon, Argon and Haiku 5.5

Google released Gemini 4 Hydrogen on November 12, 2026, priced at $1.50 per million input tokens and $7.50 per million output tokens, and Artificial Analysis measured it at $3.16 per task versus $4.97 for Gemini 4 Argon and $5.48 for Gemini 4 Carbon at current prices. Hydrogen is 25% cheaper per token than Carbon and Argon now and 62.5% cheaper from December 1, when those models double to $4.00 input and $20.00 output per million tokens, though Claude Haiku 5.5 remains 15 times cheaper per token and about 7.7 times cheaper per task at $0.41 while scoring 36 against Hydrogen's 54. Google reports Hydrogen scores 51 on the Artificial Analysis Intelligence Index (54 at its xhigh setting) and 57% on Terminal-Bench 4.1 versus 18.2% for Gemini 3.8 Flash, and its Orchestrator mode lets Carbon plan while Hydrogen executes, with Google's own Terminal-Bench 4.1 run showing near-Carbon results at 24% less quota.

Gemini 4 Hydrogen explained: why its price per token says little, and what its cost per task shows next to Carbon, Argon and Haiku 5.5 Google's new low-cost Gemini 4 model is only 25% cheaper per token than its big siblings, but it costs far less per task. We use Artificial Analysis's own run data to show why, what changes on December 1, and how Google's new Orchestrator mode splits work between Hydrogen and Carbon. This explains reporting by Google Antigravity: Gemini 4 Hydrogen in Google Antigravity November 12, 2026 https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/ . Read the original first: https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/ https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/ In one minute - Google released Gemini 4 Hydrogen on November 12. It costs $1.50 per million input tokens and $7.50 per million output tokens. - Carbon and Argon cost $2 and $10 until November 30, then $4 and $20. Hydrogen is 25% cheaper per token now and 62.5% cheaper from December 1. - Per task, Artificial Analysis measured $3.16 for Hydrogen xhigh , $4.97 for Argon and $5.48 for Carbon at today's prices. - Hydrogen is very wordy: 304 million output tokens on the index, against a median of 72.8 million. It still costs less per task. - Claude Haiku 5.5 is 15 times cheaper per token but only about 7.7 times cheaper per task $0.41 , and it scores 36 against Hydrogen's 54. - Orchestrator lets Carbon plan and Hydrogen execute. Google's own Terminal-Bench 4.1 run shows near-Carbon results with 24% less quota. What Google shipped Google is rolling out Gemini 4 Hydrogen in Antigravity, its agent coding tool, and as the low-cost member of the Gemini 4 family. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index, above Argon's 50, and 54 at its highest setting. On Terminal-Bench 4.1, a test of agent tasks in a terminal, Google reports 57% for Hydrogen against 18.2% for Gemini 3.8 Flash. Hydrogen has four thinking levels: low, medium, high the default and a new xhigh. Paid Google AI Pro and Ultra subscribers get it in Antigravity. Price per token: the confusing part Simon Willison laid out the Gemini 4 price table on November 12. Carbon and Argon each cost $2.00 in, $0.10 cached and $10.00 out per million tokens through November 30. From December 1 they double to $4.00, $0.20 and $20.00. Hydrogen is $1.50, $0.075 and $7.50. So today Hydrogen is 25% cheaper per token. From December 1 it is 62.5% cheaper: $1.50 against $4.00, and $7.50 against $20.00. Willison's post says 75%; the listed prices give 62.5%. Against the cheapest tier, Willison notes Claude Haiku 5.5 and GPT-6 Luna both cost $0.10 in and $0.50 out, so Hydrogen is 15 times their price per token. Price per task: what you actually pay Price per token is half the story. A model that thinks out loud for longer uses more tokens on every job. Artificial Analysis reports the cost of running each model through its whole index, divided per task. - Gemini 4 Hydrogen xhigh : score 54, $3.16 per task, 304M output tokens, 13.7 minutes per task. - Gemini 4 Argon High : score 50, $4.97 per task, 113M output tokens, 30.0 minutes per task. - Gemini 4 Carbon High : score 60, $5.48 per task, 205M output tokens, 28.4 minutes per task. - Claude Haiku 5.5 Max : score 36, $0.41 per task, 298M output tokens, 14.7 minutes per task. Hydrogen used the most output tokens of the four, and Artificial Analysis calls it 'very verbose'. It is still cheaper per task than Argon and Carbon. A likely reason: input, not output, made up most of its tokens 11.3 billion input against 304 million output , and Hydrogen's input price is lower too. That is our reading of the numbers, not Artificial Analysis's. Today Hydrogen is about 36% cheaper per task than Argon and 42% cheaper than Carbon. If Carbon and Argon use the same tokens after their prices double, their cost per task doubles too, to about $9.94 and $10.96. Hydrogen would then be about 68% and 71% cheaper per task. That is our arithmetic on Artificial Analysis's numbers, not a new measurement. Haiku 5.5 is 15 times cheaper per token but about 7.7 times cheaper per task, because it also writes a lot. It also scores 18 points lower on the index. How Orchestrator splits the work Orchestrator is a model option in Antigravity that pairs Hydrogen with a stronger model. Google describes it as one task that changes hands: Carbon 'can start in a read-only phase to explore dependencies, resolve ambiguities, and build a concrete plan.' Hydrogen 'could then take over to write edits, update packages, and run tests.' If Hydrogen 'hits friction', it hands back to Carbon. This is the same idea as a senior engineer writing the plan and a faster one doing the typing. The expensive model spends its tokens on judgment; the cheap one spends them on volume. Google says Orchestrator reached near-Carbon task success on Terminal-Bench 4.1, using 24% less quota on average and finishing 19% faster. That is Google's own benchmark run. Artificial Analysis does not test Orchestrator. Orchestrator Carbon and Orchestrator Argon are for Google AI Ultra subscribers; paid Pro subscribers get Orchestrator Argon . How to test it on your own work Pick five real tasks your team runs every week. Run each on your current model and on Hydrogen, at the same thinking level you would use in production. Record three numbers per run: total tokens in and out, dollars, and whether the result passed your own check. Divide dollars by passed tasks. That is your cost per task. Run the same test again after December 1, when Carbon and Argon double in price. Who is affected | Case | Status | |---|---| | Gemini API users on Carbon or Argon | Prices double on December 1; compare Hydrogen per task before then. | | Antigravity users on Google AI Pro or Ultra | Hydrogen available now; Orchestrator options depend on plan. | | Teams on Claude Haiku 5.5 or GPT-6 Luna | Still far cheaper per task; Hydrogen scores higher on the index. | What to do - Measure cost per passed task, not price per token, before you switch models. - Re-run your comparison after December 1. - If you use Antigravity, try Orchestrator on one long task and compare quota use with Carbon alone. - Watch thinking level: xhigh is what scored 54 and also what produced the heavy token use. What is still unknown - Whether Orchestrator's savings hold outside Terminal-Bench 4.1. The 24% and 19% figures are Google's own. - How Hydrogen performs per task at the default high setting. Artificial Analysis reports only xhigh. - Whether Carbon and Argon will use the same tokens after December 1. Our doubled per-task figures assume they do. - Hydrogen's price inside Google products such as Search, where Google says it will power AI features. Sources AI News Report https://theainewsreport.com/ · every headline, every morning.