Gemini 4 Hydrogen: Full Rollout, but Flash's Bargain Era Is Over
💡 One-sentence summary: Google released Gemini 4 Hydrogen on November 12, replacing 3.8 Flash as the "daily execution tier" of the Gemini 4 family. Its third-party intelligence index is 54, just one point below Claude Sonnet 5.5, and it was pushed to everyone, including free users, on launch day—but the $1.50/$7.50 pricing has no promotion, and the cheap introductory price of the Flash era is gone.
Anyone running Cursor daily or executing agent scripts has likely hit this recently: the model dropdown list changed names again.
This time it was thorough. Some developers opened Gemini and found 3.8 Flash gone, replaced by "4 Hydrogen"—no migration notice, no grace period, a direct swap. On a Hacker News thread with 480 points, that "woke up to a renamed model" confusion was among the earliest top comments.
But despite the confusion, the release—with four official blog posts in one day and an independent review agency scoring it the same day—is packed with real information. Let me walk you through the key points.
Want to try it yourself?
- 🌐 Official announcement: Introducing Gemini 4 Hydrogen
- 📊 Independent review: Artificial Analysis · Gemini 4 Hydrogen
- 🧑💻 Cursor leaderboard: CursorBench
- 💬 HN discussion: 480 points · 302 comments
What exactly was released?
First, let's get the family tree straight, because the periodic-table naming is confusing.
Google started using "elements" for the Gemini 4 family in October: Argon came first (rolled out over nearly two weeks), Carbon took over the frontier deep-reasoning tier this month, and Hydrogen, which arrived on November 12, takes the spot everyone knows as the Flash tier—the high-frequency, fast, cheap position. Google's own description is blunt: Carbon and Argon handle your hardest long-reasoning problems; Hydrogen handles fast execution for everyday coding and agentic workflows.
A few numbers worth remembering: reasoning depth has four levels—low / medium / high (default) / xhigh—with xhigh being new this time. Input natively supports 1M token context; output can also reach 1M tokens through continuation. It handles text, image, audio, video, and files.
The real news is the rollout. In the Argon era, Google's path was "security researchers first, then high-tier subscribers, then full rollout." Hydrogen is Google's first full rollout on launch day—AI Studio all tiers, Gemini API, Gemini app (free users via automatic model selection), Google Search globally, and Antigravity for Pro/Ultra subscribers all went live the same day. The Search team even gave a strong label: "the biggest Search model upgrade in history."
How strong is it? Taking official and third-party numbers together
Google's official comparison is striking: compared to 3.8 Flash, default-tier task completion is 2.5x faster, Terminal-Bench 4.1 score jumps from 18.2% to 57%, and "fabricated statements" are reduced fivefold or more.
Third-party data confirms the capability leap is real. Artificial Analysis's intelligence index (an independent composite of nine benchmarks, higher is better) gives Hydrogen 51 at high reasoning and 54 at xhigh—as reference, Gemini 4 Argon scores 50, OpenAI GPT-6.1 Terra 51, and Claude Sonnet 5.5 and GPT-6.1 Sol score 55 at their highest settings. Being surpassed by a "little brother" in the same family just four weeks after release is more telling than the score itself.
Cursor's first-party coding benchmark, CursorBench 4.0, also published a score the same day: 54.3% (default), a clear improvement over 3.8 Flash's 39.6%, landing it fifth behind Carbon High (60.4%), Claude Fable 5.2 (60.0%), Opus 5.5 (57.8%), and Sonnet 5.5 (55.5%). AlphaSignal crunched the numbers: that result comes at roughly one-tenth the price of Opus 5.5.
But two things need clarification; don't just look at the official table.
First, in the official benchmark table, Hydrogen's SWE-sweep score is 13.5%, far above Sonnet 5.5's 3.1%. Yet in the same table, Terminal-Bench 4.1 has Sonnet ahead (63.9% vs 57.0%). The two benchmarks point in opposite directions—official materials highlight the advantage, so cross-referencing is safer.
Second, pricing. Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens (including reasoning tokens), with a flat rate within the 1M window and Batch/Flex at half price—but there is no introductory promotion. Compare to 3.8 Flash's still-listed promotional price of $0.75/$3.75: output cost doubles. VentureBeat's characterization is spot-on: "a faster, smarter, pricier Flash successor."
There's another trap easy to miss: Artificial Analysis measured xhigh reasoning generating 304 million tokens on evaluation tasks, more than four times the median of 72.8 million for all models. Deeper thinking means more tokens burned—for agent pipelines billed by output, the bill may be higher than intuition suggests.
Orchestrator: Letting the model assign work to itself
There's another thing in this release that's easy to gloss over but may be the most important for agent developers: Antigravity now has Orchestrator mode.
It works by letting Carbon (planning tier) and Hydrogen (execution tier) hand off dynamically during an evolving task—the planning model passes to the execution model when it's time to act, and when the execution runs into a hard reasoning wall, it hands back. Official numbers: quota consumption down 24%, response time down 19%.
Ecosystem integration happened the same day: Cursor listed it directly, Cognition put it into Devin Desktop and CLI (FrontierCode 2.0 scored 38.2%, second only to GPT-6.1 Sol), and Mercor's APEX-Agents chart ranks it fourth.
Community reaction to this "self-routing" design is split. Supporters point to the efficiency numbers; skeptics are also loud—letting the model decide when to use which model, when models tend to overestimate their own ability, is a loop whose reliability needs time to prove. Vertical outlet LLMTracker also noted a mildly ironic detail: Hydrogen denied the existence of its own codename in conversation.
Naming, replacement, and "model fatigue"
The other side of community sentiment is clearest from the HN discussion, roughly in three buckets:
- Naming confusion: In the Pro/Flash era, the name implied the tier. After the periodic table, Argon, Carbon, Hydrogen—who's bigger than whom is all memory. A top comment sums it up: "Carbon and Hydrogen I can guess, but what is Argon?"
- Silent replacement: 3.8 Flash was swapped out directly, even Google Workspace business seats' "Everyday Help" quietly became Hydrogen. Pipelines that depend on old model behavior have to check themselves.
- Model fatigue: A representative lament on dev.to—"a new best model every few weeks, benchmarks climbing, but my daily work barely changes."
All three point to the same reality: the pace of model iteration has outrun consumers' ability to adapt.
Let's be clear
What holds up: intelligence index 54, CursorBench 54.3%—these are third-party or platform data, so the capability leap is credible; full rollout on launch day is also fact, free users can use it today.
What's debatable: no promotional pricing and doubled output cost puts real pressure on high-frequency agent users; the selective presentation of official benchmarks needs cross-checking; model self-routing reliability is unproven.
What's unconfirmed: parameter count, whether it's distilled from Carbon, the actual trigger rate of automatic model selection on the free tier—Google hasn't given any of these.
The meaning differs per audience: those running high-frequency agent pipelines should check the new bill before deciding whether to lock in low reasoning; everyday coders can try the default in Cursor directly—54.3% is worth the price; enterprise IT admins should verify whether their Workspace model has already been swapped. As for Argon—being overtaken by its own execution tier four weeks after release is probably the most telling footnote in this launch.
Sources
- Google official announcement (primary) — https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/
- Google Antigravity blog: Orchestrator (primary) — https://antigravity.google/blog/gemini-4-hydrogen-in-google-antigravity/
- VentureBeat: pricing and "pricier Flash successor" (reputable media) — https://venturebeat.com/technology/google-launches-gemini-4-hydrogen-a-faster-smarter-pricier-flash-successor
- Ars Technica: rollout strategy shift (reputable media) — https://arstechnica.com/google/2026/11/google-rolls-out-gemini-4-hydrogen-finally-replacing-3-8-flash/
- Artificial Analysis: intelligence index and cost measurements (independent review) — https://artificialanalysis.ai/models/gemini-4-hydrogen
- CursorBench leaderboard (platform benchmark) — https://cursor.com/cursorbench
- Hacker News community discussion (community) — https://news.ycombinator.com/item?id=50471431
- LLMTracker: Orchestrator community reactions (vertical media) — https://llmtracker.de/en/news/gemini-4-hydrogen-goes-orchestral-google-bets-on-self-routing-ai-and-developers-
Top 1 from juejin.cn, machine-translated. The original thread is authoritative.
Crying