Carbon put to work12 minA week of hands-on tests with Google's new frontier model, from a browser operating system to a 3D-printed wheel spacer. The model is good. The /swarm feature is what made the results look like they came from a different model entirely.Read now →

Google's Cheapest Gemini 4 Model Nearly Matches Claude Sonnet 5.5 at a Third of the Cost. Then It Got a Senior Engineer
Gemini 4 Hydrogen landed on November 12 as Google's fast, everyday model. Independent tests put it one point behind Claude Sonnet 5.5 at roughly a third of the cost per task, and a new Orchestrator mode lets it trade places with Gemini 4 Carbon mid-job.
One point. That is the gap between Google's new budget-tier model and Anthropic's mid-tier workhorse on the industry's most-watched independent scoreboard.
The price gap is much bigger. On Artificial Analysis's cost-per-task estimate, Claude Sonnet 5.5 at its maximum setting runs about $10.02 per benchmark task. Gemini 4 Hydrogen at its maximum runs about $3.16.
That's roughly a third of the money for 98% of the score. And Google didn't stop at a cheap model. It also shipped a way for that cheap model to work as the hands of its flagship, Gemini 4 Carbon, inside the same coding session.
Here's what Hydrogen actually is, what the numbers say (including the ones Google would rather you skimmed past), what builders have done with it in its first days, and what else moved in AI this week while everyone was looking at Google.
What Google launched, and when
Google announced Gemini 4 Hydrogen on November 12, 2026, with an update to the announcement on November 13. It is the third member of the Gemini 4 family, following Gemini 4 Argon and Gemini 4 Carbon, the two frontier models Google positions for deep, long-horizon reasoning.
Hydrogen has a different job. In Google's words, it is built "for rapid execution across diverse agentic workflows." Translation: it's the model you leave running all day for coding, document crunching and multi-step agent tasks, where speed and cost matter as much as raw brainpower.
The naming follows the periodic table. Argon, Carbon and now Hydrogen, the lightest element of all. For the lightweight model, it's a fair joke.
The headline specs:
- Context window: a unified window of about 1 million tokens (1,048,576), shared by input, thinking and output.
- Long output: responses can run past 64K tokens and up to 1 million tokens using background execution or client-side continuation.
- Inputs: text, images, audio and video.
- Thinking levels: four of them,
low,medium,high(the default) and a brand-newxhigh. - Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across the full context window, with cached input at $0.075 per million.
It replaces Gemini 3.8 Flash as Google's fast model. Google's developer docs even ship a migration checklist for 3.8 Flash users.
The benchmark picture, honestly
Google lined Hydrogen up against the "middle" models from its two main rivals, the tier just below each lab's top model that ends up handling a huge share of everyday work. On the Anthropic side, that's Claude Sonnet 5.5.
Start with what Google reports itself.
Where Google says Hydrogen wins
- SWE-sweep, a test where the model has to find and repair bugs in real code without breaking anything else. Scores on this one are tiny because it's brutally hard, and Google says Hydrogen outperforms similar-tier models.
- Agents' Last Exam v1.1, which checks whether a model can carry a professional workflow from start to finish: 71.6%.
- Multimodal Knowledge Agent (MKA), a new benchmark of end-to-end knowledge work over complex visual files: 53.4%, which Google says beats other models.
- HLE-Diamond, a hard reasoning set: 58.8%, which Google says leads its class.
- Vals Tax Agent Bench: #1 overall at 52.3%. Vals Legal Research Bench: tied for #1 at 55.8%.
- Agent Arena, arena.ai's live benchmark for multi-step tasks: ranked #5.
That MKA result is the one Google is pitching hardest for office work. Throw a pile of PDFs, spreadsheets, charts, audio and recorded video at Hydrogen and get back a cited report or structured deliverable.
Where it loses
Google's own materials don't hide every loss, and the independent numbers fill in the rest.
- Terminal-Bench 4.1 (agents doing real work in a command line): Hydrogen scores 57%, more than triple 3.8 Flash's 18.2%. But per VentureBeat's reading of Google's table, it still trails Claude Sonnet 5.5's 63.9%.
- Computer use: on tests where a model drives a desktop like a person clicking around apps, Hydrogen sits slightly behind Sonnet.
- Terminal-Bench-Science: on Artificial Analysis's scientific terminal tasks, Sonnet 5.5 at max scores 53.3% versus Hydrogen's 49.5% at xhigh.
So the honest summary is: Hydrogen is right there with the leading mid-tier models, ahead on most of what Google chose to test, and behind on hands-on terminal and computer-control work.
The independent scoreboard
Artificial Analysis runs every major model through the same nine-test suite (the Intelligence Index v5.0.1, covering coding, document analysis, science, finance, legal work and more) and boils it down to one number.
| Model and setting | Intelligence Index | Est. cost per task | Time per task |
|---|---|---|---|
| Claude Sonnet 5.5 (max) | 55 | $10.02 | 33.3 min |
| Gemini 4 Hydrogen (xhigh) | 54 | $3.16 | 14.8 min |
| Gemini 4 Hydrogen (high, default) | 51 | $1.87 | 8.4 min |
| Claude Sonnet 5.5 (xhigh) | 50 | $4.09 | 14.5 min |
| Gemini 4 Hydrogen (medium) | 41 | $0.80 | 3.8 min |
| Gemini 4 Hydrogen (low) | 27 | $0.29 | 1.4 min |
Three things jump out.
- At its default setting, Hydrogen (51) already beats Gemini 4 Argon (50), one of Google's own frontier models. Google put that line in its announcement for a reason.
- At comparable thinking settings, Hydrogen wins on score and cost. Hydrogen at xhigh (54, $3.16) beats Sonnet at xhigh (50, $4.09).
- Speed is real. Sonnet's top setting took 33.3 minutes per task on average in these runs. Hydrogen's top setting took 14.8.
Some subscores are striking. On AutomationBench-AA, where a model uses software tools to run automated workflows, Hydrogen at xhigh hits 76.1% to Sonnet max's 71.8%. On HLE-Diamond it's 58.8% to 53.3%. On the AA-Omniscience knowledge test, 46 to 30.
But Sonnet still wins Terminal-Bench 4.0 (63.6% to 53.5%), FrontierCode 2.0 (41.7% to 39.0%) and the AA-Briefcase test. If your day is mostly autonomous terminal work, the gap matters.
The case for "good enough"
The argument now circulating among developers, including in Chinese tech coverage, is simple: if your daily work leans on a mid-tier model from Anthropic or OpenAI, near-top quality at a fraction of the cost and well over twice the speed beats being number one on a leaderboard.
That's directionally true, with one caveat VentureBeat flagged. Hydrogen isn't cheap compared with true budget models. Per million tokens, it costs about 15 times the lowest standard rates for Claude Haiku 5.5 and OpenAI's GPT-6 Luna. Its sweet spot is replacing mid-tier spend, not bargain-bin spend.
Speed and the four thinking levels
Google's single biggest pitch is speed. At its default high setting, Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash. When you're running checks on your code over and over, that adds up fast.
The thinking levels let you choose how hard it works:
- Low: fastest and cheapest. Good for quick lookups and simple edits.
- Medium: a real step up in quality at modest cost.
- High (default): the balance point most people will use.
- Xhigh: new with Hydrogen, for tasks that need the deepest reasoning.
Google says quality scales "in a smooth, predictable progression," so you only pay for the depth a task needs.
One useful nuance from Artificial Analysis's numbers: on Terminal-Bench 4.0, xhigh bought no improvement over high (both about 54%), while the full-suite cost rose from $1.87 to $3.16. Don't reflexively crank the dial.
Google also claims Hydrogen cuts fabricated claims by more than 5x versus 3.8 Flash and adapts cleanly when a tool call fails or a search returns nothing. Partners in early testing reported 2x to 2.8x faster runs with 12% to 23% fewer tokens on data-heavy enterprise work.
Box offered an outside data point. CEO Aaron Levie posted that Box's agent went from 74% to 79% accuracy on its Complex Work Eval moving from Flash to Hydrogen, in less than half the time and with about 12% fewer tokens.
What people built in the first 72 hours
Benchmarks are abstract. Demos aren't.
A walkable 3D island from one prompt
Google's flagship demo is Geminisland, a Minecraft-style shared voxel island. You type a prompt, and Hydrogen codes a custom structure on the spot, placing it live on a map players can walk around in first person. Roads, buildings and oversized animals all appear on demand.
A canyon flight game
One user on X built a canyon flight game with Three.js, the popular library for 3D graphics in the browser. It had code-generated landscape, moving water and 25 scoring gates to fly through.
Light, bent in real time
Google's water-and-glass visualizer is built "from pure math" and simulates light refracting, reflecting, splitting into colors (dispersion) and being absorbed as it passes through water and glass, live. That takes graphics code plus a genuine grasp of optics.
Ships in the glaciers: Hydrogen vs Carbon
One developer gave the same job, a 3D scene of ships sailing through glaciers, to both Hydrogen and Gemini 4 Carbon. By that builder's account, Hydrogen took 1 hour 23 minutes and about $4.27. Carbon took 2 hours 39 minutes and about $8.83. Roughly half the time, half the money, and a result that looked close to Google's flagship.
The pelican test
A long-running informal AI test is asking a model to draw a pelican riding a bicycle as an SVG, an image written in code. One user made it much harder: a fully animated pelican riding through five changing seasons. They report Anthropic's top model, Claude Fable 5.2, took about 25 minutes. Hydrogen took about five, with a solid-looking result.
One anecdote is not a benchmark. But it matches the pattern in the independent timing data.
An entire edited video
The demo that will make video editors wince: using a custom video-assembly skill and a creative brief, Hydrogen produced an archival sizzle reel about the element hydrogen. It sourced public-domain footage, cut it into one continuous edit and added narration.
Office work, everywhere
Google also showed Hydrogen:
- Building a one-shot e-commerce storefront with a WebGL2 shader and a live product configurator.
- Turning Google Sheets into an animated pixel canvas using formulas and conditional formatting.
- Generating a full Google Slides deck from a keynote video, then auditing and fixing its own layout errors.
- Building interactive financial models in Sheets more than 30% faster than earlier models.
- Mapping real-time lightning strikes onto a 3D globe that plays ambient music.
Orchestrator: the senior engineer and the fast junior
This is the part of the launch with the longest shelf life.
When Carbon launched, it came with Swarm, a feature in Google's AI coding app, Antigravity, that throws entire teams of agents at huge engineering jobs in parallel. Google has said tens of thousands of swarms were launched within a couple of days.
But for everyday work, one feature or one bug fix, running your top model with a squad of agents on every step is slow and burns through your usage limit.
Why routers aren't enough
The usual fix is a model router. It reads your first message, guesses how hard the job is, and sends the whole thing to the cheapest model that can handle it.
The problem: agents now run for a long time, and a job that looks easy can turn messy halfway through. A router makes one decision up front and lives with it.
How Orchestrator works
Orchestrator lets Carbon and Hydrogen switch places within the same session:
- Carbon goes first, read-only. It explores the codebase without changing anything, works out how the pieces connect, resolves ambiguities and writes a concrete plan with checkpoints for rethinking.
- Hydrogen does the hands-on work. It writes the edits, updates dependencies and runs the tests.
- If Hydrogen hits a wall, Carbon steps back in. It reviews what happened and repairs the plan.
Carbon is the senior engineer. Hydrogen is the fast junior who does the typing. For parallel jobs, Carbon can coordinate while Hydrogen powers the worker and verifier agents that build and test changes across isolated workspaces.
The numbers
On Terminal-Bench 4.0, Google says Orchestrator reached near-Carbon task success while using 24% less quota and finishing 19% faster on average. Hydrogen alone scores 57% on Terminal-Bench 4.1, more than triple 3.8 Flash's 18.2%.
Developers can wire this up themselves with the Antigravity SDK. In Google's demo, Hydrogen agents keep pace with a flood of bug reports on an open-source project, searching code, asking clarifying questions and making quick design tweaks from review comments. A Carbon agent takes the bigger rewrites, new features and security checks.
Who gets what
- Gemini app and Google Search: rolling out to everyone. Free Gemini app users get Hydrogen through automatic model selection when it fits the prompt. Google calls it the biggest model upgrade to Search yet.
- Developers: Gemini API and AI Studio on both free and paid tiers, plus Android Studio. Hydrogen is now the default model for the Antigravity managed agent.
- Antigravity: paid Google AI Pro and Ultra subscribers.
- Orchestrator: comes in a Carbon version and an Argon version. Ultra gets both. Pro gets the Argon version.
- Businesses: Gemini for business, Antigravity and Gemini in Sheets for eligible plans.
Why this matters
The middle tier is where the money is. Frontier models get headlines, but mid-tier models handle the bulk of real work. A model that nearly matches Sonnet 5.5 at a third of the cost per task resets what teams should expect to pay.
Multi-model is becoming the default. Orchestrator is an admission that no single model is right for an entire job. Expect every lab to ship its own planner-plus-executor pairing.
Search just changed for billions of people. Hydrogen is now inside Google Search for all users, with better handling of messy, multi-part questions and a wider "query fan-out" that Google says surfaces more links. If you publish online, that's the model now reading and summarizing your pages.
Practical takeaways
- If you pay for a mid-tier model, run your own bake-off. Pick 20 real tasks from your week and run them on Hydrogen at
high. Compare quality, time and bill. - Start at
high, notxhigh. Move up only when you see failures. - Keep Sonnet-class models for terminal-heavy agents until Hydrogen closes that gap, or let Orchestrator hand the hard parts to Carbon.
- Migrating from 3.8 Flash? Read the checklist. Hydrogen defaults to high reasoning (Flash defaulted to medium) and rejects the old
thinking_budgetparameter andMINIMALsetting. - Watch caching. On Google Cloud, automatic caching starts at 4,096 tokens, and explicit caching wasn't available at launch.
Counterpoints and risks
- Vendor benchmarks are curated. Google chose its comparison set. The independent index is closer, and it still shows Sonnet ahead on terminal and coding-frontier tests.
- Cost per task is an estimate. Artificial Analysis weights its own suite. Your workload's token mix may differ.
- Speed claims vary by load. Early-days numbers can shift as demand grows.
- Demos are best cases. The glacier and pelican comparisons are single runs reported by individual users.
- Lock-in. Orchestrator only works inside Google's stack. Cheaper per task can still mean more dependent.
Also this week: Microsoft teases, OpenAI lets agents pay
Microsoft: an announcement of an announcement
Microsoft AI said its in-house MAI team will bring "something new" to Microsoft Ignite, which starts November 17: new models, demos and speaker sessions from the people who built them. Mustafa Suleyman, who runs Microsoft AI, reposted it. No model names, no demo titles, no speakers.
So, for now, Microsoft has announced that it will announce something. The rest of the Ignite agenda leans heavily on agents, with Copilot Studio for building them and Agent 365 for securing and monitoring them.
ChatGPT can now spend your money
On November 13, OpenAI launched ChatGPT Wallet, built with American Express, Mastercard and Visa. It lets ChatGPT's agents, called dots, help you choose what to buy and then check out.
How a purchase works:
- You add an eligible card once.
- You ask your dot to help you buy something.
- Before checkout, you see the items, the total and the card.
- You approve with Face ID or a fingerprint, which creates a record of what you authorized.
- The payment network issues a payment token, a stand-in card number, so your real card details stay hidden from both the agent and the store.
The store still gets your shipping address and contact details, and you keep your card's rewards and dispute protections. For now it's limited to eligible US ChatGPT Pro subscribers, 18 and over, with US-issued cards. OpenAI says it eventually wants users to set spending limits, purchase categories or time windows so a dot can buy pre-approved things on its own.
The catch is stores. Some retailers block AI agents. The biggest of them, Amazon, has fought outside shopping agents (it sued Perplexity in 2025 over its Comet browser agent) while promoting its own assistant. To win merchants over, OpenAI published commerce guidance on making sure agents aren't accidentally blocked.
FAQ
Is Gemini 4 Hydrogen free? Free Gemini app users can get it through automatic model selection, and it's in Google Search for everyone. The API has a free tier in AI Studio; paid usage is $1.50/$7.50 per million input/output tokens.
Is Hydrogen better than Claude Sonnet 5.5? Not overall. Sonnet 5.5 at max scores 55 on the Artificial Analysis index to Hydrogen's 54 at xhigh, and leads on terminal and computer-use tasks. Hydrogen is far cheaper and faster per task, and wins on several knowledge-work and automation tests.
What is Orchestrator? An Antigravity feature that switches between Gemini 4 Carbon (planning, review) and Hydrogen (execution) in a single session. Google reports near-Carbon success with 24% less quota and 19% faster runs.
Which thinking level should I use?
Start with the default, high. Use xhigh only for tasks that fail at high, since it can cost noticeably more without always scoring higher.
Does Hydrogen replace Gemini 3.8 Flash? It's Google's successor for fast workloads, with a migration guide. Google's deprecation schedule lists no shutdown date for Flash yet.
The bottom line
Gemini 4 Hydrogen isn't the smartest model on the market. It doesn't need to be.
It lands one point behind Claude Sonnet 5.5 on the toughest independent scoreboard, at roughly a third of the cost per task and in under half the time. And with Orchestrator, Google is betting the future isn't one giant model doing everything but a senior model that plans and a fast one that types.
If you're paying mid-tier prices for everyday AI work, this is the week to rerun your math.
Sources
- Introducing Gemini 4 Hydrogen (Google)
- What's new in Gemini 4 Hydrogen (Google AI Studio docs)
- Gemini 4 Hydrogen vs Claude Sonnet 5.5 (Artificial Analysis)
- Google launches Gemini 4 Hydrogen (VentureBeat)
- Introducing ChatGPT Wallet (OpenAI)
- ChatGPT's new Wallet lets AI agents shop and check out for you (TechCrunch)





