Gemini 4 Hydrogen Replaced Flash and Doubled the Price
For two weeks after Google shipped Gemini 4 Carbon, people online kept joking that the cheap one would have to be called Hydrogen. On November 12 Google shipped it, and it’s called Hydrogen. It replaced Gemini 3.8 Flash in the Gemini app without much of an announcement. One Hacker News commenter only found out when the model picker said “4 Hydrogen” and they had to go look it up.
If you pay for API calls, the first number to know is $7.50. That’s Hydrogen’s price per million output tokens, double the $3.75 that 3.8 Flash charges. It’s also a much smarter model. Two days before Hydrogen landed, Carbon left its cyber-defenders-only preview and went public, so every model leading an Arena category this week is one you can actually buy.
The other date to know is November 30. That’s when Google’s intro pricing on Argon and Carbon ends. My cheapskate table didn’t change a single pick this week, and every Gemini-led ratio in it still doubles on December 1.
- Gemini 4 Hydrogen: the Flash replacement that isn’t Flash-priced
- Gemini 4 Carbon: the gated #1 goes public
- The cheap end: Hy4 Flash, a free ERNIE, and Qwen4-Max
- The rest of the board
- The cheapskate table
- Local sweet spot
- Horror stories
- What’s coming
- Finally
Gemini 4 Hydrogen: the Flash replacement that isn’t Flash-priced
Google pitches Hydrogen as the fast member of the Gemini 4 family, for everyday coding and agent work, while Argon and Carbon handle the slower, deeper reasoning. The API price is $1.50 in and $7.50 out per million tokens, with a flex tier at half that. It’s already serving Search and the Gemini app, and it showed up in Google’s Antigravity IDE with a new feature called Orchestrator.
On Artificial Analysis, the score depends heavily on how hard you let it think:
| Hydrogen effort | AA Index |
|---|---|
| xhigh | 54 |
| high | 51 |
| medium | 41 |
| low | 27 |
Gemini 3.8 Flash tops out at 31. So yes, Hydrogen costs double per token, but at high effort it’s a different tier of model. At xhigh it’s seventh on AA’s whole index, ahead of Argon’s 50, at $3.16 per AA task. AA does warn that it’s “very verbose” and slower than average per task. It generates around 174 tokens a second, but the first answer token takes about 45 seconds at xhigh, so it’s fast at writing and slow to start.
The Arena board published November 11, a day before launch, already lists Hydrogen, so people were voting on it before anyone knew its name. It sits 15th Overall at 1490.43 on 4,741 votes, sixth in Creative Writing, sixth in Instruction Following, 15th in Hard Prompts. That puts it inside five of my six bands. It’s not the cheapest model in any of them, though, and it missed the Coding band by 0.48 points.
The launch thread hit 572 points. The Antigravity Orchestrator thread got another 507, and one comment there sums up the Gemini 4 family so far: “Argon and Carbon are very slow. In the time it takes for them to complete a single task, Opus already managed…” So a fast Gemini 4 is what people wanted. Whether it’s worth twice Flash per token depends on whether your work ever needed more than 31 points of brain.
Gemini 4 Carbon: the gated #1 goes public
Last week Carbon led three Arena categories, and only cybersecurity defenders in Google’s Fairwind program could call it. On November 10 Google rolled it out through the Gemini API, AI Studio, Gemini for business, the Gemini app, AI Mode in Search and Antigravity for Ultra subscribers. The intro price is $2/$10 through November 30. After that it’s $4/$20, the same schedule as Argon.
It’s live on OpenRouter too, and it put up 1.02 trillion tokens in its first week, third on the New & Trending list. AA has it at 60 on high effort, second only to Claude Fable 5.2’s router-assisted 65. AA’s cost and latency numbers for Carbon also moved: $5.48 per task and a 72.56-second first answer, down from $7.76 and 147.83 seconds in the preview measurements. I wouldn’t read much into that until AA says what changed.
Two notes from the GA thread. One person ran an unfinished paper draft past several frontier models, a test they’ve been running for a while, and said Carbon was “surprisingly good” compared with Astra, Fable, Opus and Argon. Another opened the Gemini app on a Pro plan and still only had 3.1 Pro and 3.6 Flash, then wrote that this would be their last month. The API is open to everyone. The consumer app is still rolling out by plan tier.
Because Carbon is purchasable now, I’m retiring the “minus the model you can’t call” shadow table from the last two issues. Every leader in this week’s table is something you can put on a credit card.
The cheap end: Hy4 Flash, a free ERNIE, and Qwen4-Max
Hy4 Flash
Tencent put Hy4 Flash on OpenRouter on November 16 at $0.035 in and $0.14 out, which is a penny above Qwen3.7 Flash for the bottom of the market. It did 1.88 trillion tokens on its first day and landed third on the daily board behind the two DeepSeek V4.1 Flash snapshots. The parent model, Hy4, is second for the week at 16 trillion, flat.
It has no Arena rating and no AA score yet. Tencent is the only provider. One trillion-plus token day at fourteen cents an output million tells you people are curious and the price is low. It doesn’t tell you whether the model is any good. The full Hy4 weights that Tencent called “coming days” on November 4 still haven’t shipped, as far as I can find.
ERNIE 5.5 Flash Preview, free
Baidu’s ERNIE 5.5 Flash Preview is free on OpenRouter with a 1M context window. It topped the New & Trending list at 2.47 trillion tokens, which is a lot of people grabbing free inference while it lasts. One provider, Baidu Qianfan, and the commercial terms aren’t clear. We’ve watched this happen with Step 5 and MiniMax M3.1 Flash: a free model shoots up the board, the promo ends, and the traffic goes looking for the next free thing.
Qwen4-Max
Alibaba opened Qwen4-Max as a preview on November 11, through its Token Plan subscription. According to an aggregator summary of Alibaba’s notice, that preview runs until November 25 and then retires. Even so, OpenRouter already lists qwen/qwen4-max at $1.425/$4.275 from Alibaba, created November 16. I can’t tell yet which of those two is the version that sticks around.
The blog post thread got 866 points, the biggest model launch on HN this week. The comparison worth having came from one commenter: Qwen4-Max scores 43 on AA at $3.27 a task, while GPT-6.1 Sol at medium gets 46 for $0.36. That’s nine times the cost per task for a lower score. It also has no Arena rating yet. One more thread to watch: someone asked whether Alibaba removed reverse-engineering and offensive-cyber capability, because their RE test prompts were coming back as generic connection errors. That’s unconfirmed. If it’s true, Qwen joins Kimi K3.1 Pro and Hy4 on the list of models shipped with the cyber skills taken out.
The rest of the board
OpenRouter data runs through November 16. DeepSeek V4.1 Flash 1017 is first for the fifth week straight at 61.2 trillion tokens, up 10 percent. That’s 209 trillion over thirty days. DeepSeek’s share of text requests held at 31.8 percent. Arena still hasn’t rated the 1017 snapshot, five weeks after it took the top spot.
Behind it: Hy4 at 16 trillion, the original V4.1 Flash at 15.5 trillion (down 12 percent), and GLM-5.3-Flash at 11.8 trillion (down 3). MiMo-V2.6-Flash jumped 63 percent to 4.16 trillion. Google’s request count rose 14 percent, which makes sense with two new Gemini 4 models landing in one week.
On spend rather than volume, Claude Opus 5.5 takes 25.6 percent of OpenRouter’s implementation-task spend. Kimi K3.1 Pro came out of nowhere to second place at 7.9 percent, a week after its weights hit Hugging Face.
The cheapskate table
How the picks work, if this is your first issue. Each Arena category gets a line 50 points under its leader, and the pick is whichever model above that line has the lowest output price. A script handles the filtering across all ~420 rows of LMArena’s dataset (this week’s board is dated November 11) and records the first model under the line, so I know no band was cut off early. Prices come from each maker’s own OpenRouter endpoint. Argon and Carbon are at their $10 intro rate.
| Category | Leader | $ out | Cheapskate pick | $ out | Δ rating | Cheaper by |
|---|---|---|---|---|---|---|
| Overall | Gemini 4 Carbon (High) (1526) | $10 | MiMo-V2.6-Pro (1480, #31) | $0.87 | −46 | 11.5× |
| Coding | Gemini 4 Carbon (High) (1568) | $10 | GLM-5.3-Flash (1526, #28) | $0.50 | −42 | 20× |
| Creative Writing | Gemini 4 Argon (High) (1523) | $10 | Gemini 3.8 Flash (High) (1488, #10) | $3.75 | −34 | 2.7× |
| Instruction Following | Gemini 4 Argon (High) (1528) | $10 | Gemini 3.8 Flash (High) (1491, #15) | $3.75 | −36 | 2.7× |
| Hard Prompts | Gemini 4 Carbon (High) (1547) | $10 | GLM-5.3-Flash (1500, #35) | $0.50 | −46 | 20× |
| Math | Gemini 4 Argon (High) (1536) | $10 | GLM-5.3-Flash (1505, #11) | $0.50 | −32 | 20× |
Band depths: Overall 32, Coding 41, Creative 16, Instruction Following 28, Hard Prompts 41, Math 31.
Nothing moved, and some of it was close
Same six picks as last week, and on most lines that’s the third week running. Three of the six still belong to GLM-5.3-Flash at fifty cents (Coding, Hard Prompts, Math). MiMo-V2.6-Pro’s grip on Overall got looser: it clears the line by 3.71 now, down from 4.09. Creative and Instruction Following stay with Gemini 3.8 Flash, which beats its 3.7 sibling on rating at the same $3.75.
The edges were tight. Overall’s first model below the line is GPT-5.5, 0.06 points under. In Coding it’s GPT-5.5 (High), 0.08 under. In Math, Muse Spark 1.2 missed by 0.02. I compute this to the decimal and don’t round, so a point or two of drift next week could change who’s in. GLM-5.3-Flash is out of Overall by 0.92 for the third straight week.
What you give up with the two cheap Chinese picks is still time. On AA’s agent tasks GLM-5.3-Flash needs about 40 minutes a task and MiMo-V2.6-Pro about 37, and each of those tasks costs $0.58 and $0.26 respectively, against Carbon’s $5.48.
Creative Writing and Instruction Following still have the smallest discount. The cheapest model in either band costs $3.75, and Hydrogen rates above 3.8 Flash in both for twice the price. If you’re writing prose and the budget has room, Hydrogen is the step up I’d test first.
The December 1 cliff
Argon and Carbon go from $10 to $20 output on December 1. No pick has to move for this to change the table:
| Category | Cheaper by today | Cheaper by on Dec 1 |
|---|---|---|
| Overall | 11.5× | 23× |
| Coding | 20× | 40× |
| Creative Writing | 2.7× | 5.3× |
| Instruction Following | 2.7× | 5.3× |
| Hard Prompts | 20× | 40× |
| Math | 20× | 40× |
If you’re building on Argon or Carbon right now because the intro price made it easy, you have two weeks to decide whether the work justifies double. Nothing I found mentions an intro period on Hydrogen’s $7.50, so after December 1 it may be the Gemini 4 model you can afford.
Local sweet spot
Each row is the largest model that fits on that machine with memory to spare for context and still manages 15 or more tokens a second.
| Your machine | Pick | Size | Arena | AA | Measured tok/s |
|---|---|---|---|---|---|
| Mac 16GB | qwen3.5-9b Q4_K_M | 5.7 GB | unlisted | - | 27 (n=4, 6-82) |
| Mac 32GB | qwen3.8-27b MLX 4bit | 16.1 GB | 1434 (#111) | 34 | not measured |
| Mac 64GB+ | gemma-4-31b Q4_K_M | 18.3 GB | 1452 (#79) | 15 | not measured |
| NVIDIA 12GB | qwen3.5-9b Q4_K_M | 5.7 GB | unlisted | - | 50 (n=1) |
| NVIDIA 24GB | gemma-4-31b Q4_K_M | 18.3 GB | 1452 (#79) | 15 | not measured |
No changes. The two new small open models on Arena, Inkling 2 Small (1411.78) and Muse Glimmer (1424.64), both rate under Gemma 4 31B. The 9B rows are still picked by what fits, since nothing that size has an Arena rating, and the AA column uses the older v4.3.2 scale. The “not measured” cells are still waiting on someone to post real numbers.
Horror stories
Reddit blocked my research tool again this week, so everything here comes from Hacker News. If the list looks short, that’s why. It doesn’t mean nothing went wrong.
Claude went down and took Claude Code with it
On November 13, Anthropic’s status page posted “Elevated errors on claude.ai, Claude Code, and platform.claude.com.” The HN thread reached 493 points. One user said their sessions suddenly demanded a re-login, which worked once and then failed with “OAuth request failed… Temporarily unable to authenticate.” Others noticed Microsoft 365 was having problems at the same time. It was fixed in roughly two hours. One commenter joked that nobody remembered the git command to push anymore. If your whole workflow runs through one vendor’s agent, two hours of downtime means two hours of not working.
Opus 5.5 still argues with the ticket
Same thread, a smaller complaint. A developer back on Claude after a stretch with Codex said Opus 5.5 “disregards specifications in tickets if it thinks it knows better” and “complains about the size of tasks” (smallerfish). That’s one person’s experience. But Opus 5.5 is the most-paid-for model on OpenRouter’s implementation work, so it’s worth knowing that someone thinks it doesn’t follow instructions.
Humble Bundle’s login, allegedly vibecoded
A forum post claiming Humble Bundle “killed most of their security through vibecoding” got 415 points. The claim is that a site rewrite dropped 2FA. Before you repeat it, read the replies: several commenters point out there’s no actual proof the rewrite was AI-written. One did report that logging in failed for them through both Google and email. So something broke in the rewrite. Whether AI wrote it is still a guess.
What’s coming
On November 25 the Qwen4-Max Preview on Token Plan is scheduled to retire, and we’ll find out whether the OpenRouter listing stays. November 30 is the last day of intro pricing on Argon and Carbon before they go to $4/$20.
Hy4’s weights are still “coming days” after two weeks, and the ERNIE 5.5 Flash weights are promised under Apache 2.0 by one source. DeepSeek V4.1 Pro hasn’t shipped either: a November 13 check of DeepSeek’s changelog found nothing new.
Last week I passed along a rumor list from one X account. Both GPT-6.2 windows (Astra and Sol) closed on November 15 with nothing shipped. DeepSeek V4.1 Pro didn’t show up around the 12th. Qwen 4 Max did arrive, sort of, as a preview on the 11th. Grok 4.8 still has until November 30. That’s one partial hit out of five, with one still open.
On Arena I’m waiting for first ratings on Fable 5.2, Hy4, Hy4 Flash, Qwen4-Max and DeepSeek 1017. Hydrogen’s Coding rating could move into the band with a few more votes. Polymarket already has markets on a “Gemini Krypton.” Google really is working through the periodic table.
Finally
If you’re using Gemini 3.8 Flash and it’s working, Hydrogen isn’t an automatic upgrade. You’d pay double the output rate and wait longer for the first token, and that only pays off on work that needs the extra 20-odd AA points. If you’re on Argon or Carbon at the intro price, put November 30 in your calendar. And if you’re mostly trying to keep the bill down, the table is the same as last week: GLM-5.3-Flash at fifty cents for coding, hard prompts and math, and MiMo-V2.6-Pro at 87 cents for everything else, as long as you don’t mind waiting.
Next week the Qwen preview window closes, the Gemini price deadline gets closer, and Hy4’s weights get one more week to actually show up.
