
Gemini 4 Hydrogen Is the Model Google Handed Your Inbox: What the Free Tasks Rollout Actually Runs On
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0711 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 86 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 124 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 43 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 · 192 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 130 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 135 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 182 tok/s
On 17 November 2026 Google switched on Tasks for free Gemini users in six countries, and the model behind it is not the one the company spent October promoting. Tasks runs on Gemini 4 Hydrogen — the $1.50-per-million-input, $7.50-per-million-output model that went generally available on 12 November — not on Gemini 4 Carbon, the $2/$10 flagship that leads Google's own line on the Artificial Analysis Intelligence Index. That choice is the story. An agent that reads your mail, holds a payment card and phones car dealerships on your behalf now runs on the cheap tier, and Google published a page of safety numbers arguing that this is fine. Whether it is fine is the question a reader should leave with.
What shipped on 17 November, precisely
The announcement is Josh Woodward's, Google's vice president of Labs, Gemini app and AI Studio, posted to the company blog under the headline "Tasks in Gemini are now free of charge, with new ways to get things done on your behalf." Sundar Pichai's post on X went out the same day at 17:20 UTC and led with the framing the product team has settled on: knowing things, then doing things, then finding ways to save you time and money. The geographic claim in both is identical — hundreds of millions of people across the U.S., Japan, Korea, Canada, Australia and India — which is six countries, and the same six in each telling.
What you can actually do, according to Google's own instructions:
• Flip the home-screen toggle in the Gemini app from Chat to Task, then describe the job in plain language.
• Forward a batch of emails to [email protected] and have the dates land on your calendar with other people invited. Google's example is a pile of school emails, resolved in under a minute, with a carpool conflict spotted and a text offered to the other parent.
• Text Gemini over RCS on Android or iOS — but that path requires a U.S. phone number on an eligible carrier, and it is set up through Settings, then Channels, then "Text with Gemini."
The advanced features are narrower than the headline. Calling businesses, paying bills through the Agent Payments Protocol, checking out with Google Pay against the Shopping Graph, and the Google Password Manager integration that lets Gemini read a utility portal are all described as rolling out "starting in the U.S." Google's footnote list adds two constraints most summaries skip: calling through Gemini is not available in Indiana, Louisiana, Minnesota, Montana or Nebraska, and it needs a U.S. number in the first place.

The word "free" is doing a lot of work
Google's headline says Tasks is "now free of charge," which tells you the previous state without documenting it. The post does not say what early access cost, who had it, or what happens to a subscriber who was already paying for it. That is worth flagging plainly rather than assuming the most flattering reading.
What is also not published is a quota. There is no stated number of tasks per day, no stated token ceiling, no stated limit on how many emails you can forward. For a feature whose whole pitch is "hand off your life admin," the absence of a published ceiling is the single most important thing missing from the announcement, because the cost model underneath it is metered on every token the model produces — and, as the numbers below show, this particular model produces a lot of them.
Why Hydrogen rather than Carbon
Google's own explanation is structural. In the 12 November launch post, Koray Kavukcuoglu's team positions Carbon and Argon as the models for "deep, long-horizon reasoning on your hardest problems," and Hydrogen as the one "built for rapid execution across diverse agentic workflows." A task like booking a service appointment is not a reasoning problem; it is a sequence of tool calls where latency is felt by a human waiting and a wrong click costs money. That is the profile Hydrogen was tuned for, and it is cheaper to run at volume.
The safety page is where the argument gets specific, and every figure below is vendor-reported and unreproduced outside Google:
• Overreach in agentic tasks is down by up to 81% against Gemini 3.8 Flash; under-service is down by up to 79%.
• Misaligned actions and task shortcutting across tool-use and coding evaluations are cut 12-fold.
• On impossible or blocked tasks, the model is up to 2x less likely to bypass permission boundaries, and scores 94% on Google's internal agent-honesty evaluation.
• Against indirect prompt injection on the Gray Swan IPI benchmark it is more than 2.3x more resilient, and Google claims 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns.
• False refusals on HealthBench Pro stay under 0.4%, which matters for a product that will be asked about medical bills.
Those claims are the licence for the rollout. The architecture around them is the part you can verify by using it. Each task runs inside Google Cloud's Agent Compute Engine — a hardware-isolated virtual machine dedicated to one account and never shared — with prompt-injection filters scanning search results and pages, Google Safe Browsing limiting which links the agent may open, and a designed pause for confirmation before anything like a purchase. Google also says the agent will ask clarifying questions when an instruction is ambiguous, and that routine steps can be marked "Always allow."
The independent numbers, including the one Google would rather you skipped
Artificial Analysis publishes Gemini 4 Hydrogen on its Intelligence Index v5.0.1 at 54 in the xhigh thinking setting, ranked 23rd of 91 models on that board. For context measured the same day: Gemini 4 Carbon sits at 60, Gemini 4 Argon at 50, and the model Hydrogen replaces in the line-up, Gemini 3.8 Flash, at 31. So Hydrogen clears the model it supersedes by 23 points and clears Argon by four.

One asymmetry is worth stating because it is easy to misread from the leaderboard: the 54 belongs to Hydrogen's xhigh thinking setting, which is not the default. Google's own launch post puts the default-high figure at 51 on the same index, against Argon's 50; the 54 is what you get by paying for the most expensive setting. Carbon's headline 60 is measured at high. Comparing the two headline numbers therefore compares Hydrogen at its maximum reasoning effort against Carbon at its default.
Speed is where Google's launch claim and the harness part company. Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default. On Artificial Analysis's own task suite Hydrogen takes 14.9 minutes per index task and Gemini 3.8 Flash 14.1 — a rounding-error difference in the wrong direction. The 2.5x does hold where Google actually needs it to: against Argon at 29.2 minutes and Carbon at 33.4. The honest reading is that Hydrogen is fast for a reasoning model and roughly level with the Flash it replaces, which is a different and less quotable claim than the one in the press release.
Cost per finished task is the number that justifies the rollout, and it is better than the per-token price suggests. Artificial Analysis puts a Hydrogen index task at $3.16 against $7.77 for Gemini 3.8 Flash, $4.97 for Argon and $5.48 for Carbon. Output speed is 172 tokens per second, nearly three times Carbon's 61.
Then the caveat. Hydrogen is the second most verbose model on Artificial Analysis's 90-model board — 89th of 90 — producing 304 million output tokens across the index run, against 205 million for Carbon, 126 million for 3.8 Flash and a median of 65.4 million. It generates roughly 4.6 times the median model's output to reach its score. For a fixed-price subscription like Tasks that is Google's problem. For anyone metering the same model through the API, it is yours: a task that reasons 4.6 times as long costs 4.6 times as much at $7.50 per million output tokens, and the cheaper per-token rate only partly covers it.

If you want to run this model yourself
Gemini 4 Hydrogen is available on both the free and paid tiers of the Gemini API and in AI Studio under the model string gemini-4-hydrogen, at a flat $1.50 input and $7.50 output across the entire 1,048,576-token context window. Context caching is $0.075 per million and implicit only — explicit caching is not supported, which rules out the usual trick of pinning a long document as a reusable prefix. Batch and Flex tiers halve everything to $0.75 and $3.75; Priority doubles it to $2.70 and $13.50. Unlike Gemini 4 Carbon, whose $2/$10 introductory rate has a printed expiry of 30 November 2026, no end date appears on Hydrogen's card.
What it will and will not do, per the model card:
• Accepts text, images, audio, video and PDF; outputs text only. No image generation, no audio generation, no Live API.
• Supports function calling, code execution, computer use, structured outputs, URL context, and grounding against both Google Search and Google Maps.
• Four thinking levels — low, medium, high (default) and a new xhigh — with 65,536 tokens of default synchronous output and up to the full million through long decode.
For a developer building the same class of agent Tasks demonstrates, the useful pattern is not a single model but two: a fast executor doing the tool calls and a frontier model adjudicating when the executor gets stuck. Google's own Orchestrator in Antigravity is that idea shipped as a product, and it reports near-Carbon task success at 24% less quota and 19% faster on Terminal-Bench 4.0 — vendor numbers, on the vendor's own product.
That pattern is also what a routing layer is for, and it is worth being precise about where we sit. Gemini 4 Hydrogen is not in the OrcaRouter catalogue today; calling it means Google's own API or AI Studio. Gemini 4 Carbon and Gemini 4 Argon are, at provider list price with 0% markup passed through, so the split above — Hydrogen doing the work, Carbon stepping in when it stalls — is something you can wire to one key and one billing line rather than two contracts. If a model you route stops answering, ours fails over instead of failing the request, which is the difference between a p95 you can defend and an incident.
What to watch
Three things will settle whether this rollout was a good idea, and none of them are visible yet. The first is a quota: the moment Google publishes a daily limit on Tasks, the economics of running an agent on a 4.6-times-median-verbosity model become public arithmetic. The second is the Universal Commerce Protocol integration Google says is coming, which moves the agent from "asks before it buys" to a checkout path it owns end to end. The third is the failure report. Google's own post concedes that some sites let the agent get most of the way and then hand the screen back, and that tasks will sometimes miss — with a thumbs-down as the only remedy offered. Whether that feedback loop tightens the tool-call reliability fast enough is the part of this that a model card cannot tell you.
For now the practical takeaway is narrow and useful. Tasks is live and free in six countries as of 17 November, on a model that costs $1.50 and $7.50 per million tokens, that is fast rather than clever, that will occasionally talk four times longer than it needs to, and that Google has bolted to a sandbox, an injection filter and a confirmation prompt because it is going to be holding your card. If you are building anything similar, those four facts are the specification.
