Quick answer
On ten small coding tasks, Gemini 4 Argon did not beat Gemini 4 Hydrogen: both passed almost everything at every level we ran, and Argon costs more. Hydrogen passed 28 of its 29 graded cells at Low, Medium and High; Argon, run in Antigravity, passed 30 of 30. Argon’s list price is 1.33× Hydrogen’s per token today and 2.67× from December 1, and on the benchmark table we read, Argon at High costs $4.97 a task to Hydrogen’s $1.87 for an index score one point lower. Tasks this size do not show what Argon is for; its model card points at long agentic work and very long context.
Six things to take
- Same tasks, no winner. Both models passed nearly every task at every level. Small, checkable tasks do not separate them.
- Per token, Argon costs a third more today and more than double in December. $2.00 and $10.00 per million tokens against Hydrogen’s $1.50 and $7.50, rising to $4.00 and $20.00 on December 1.
- The same model differed between apps. Argon at Low missed the CSV task in AI Studio and passed it in Antigravity. One run per cell shows variation, not a rate.
- At Low, Argon was a little slower. Median 7.6 s a task against Hydrogen’s 5.3 s in AI Studio.
- Carbon is not open to test for free. We cover it from its card and the price page only, labelled as read.
- The free limits are the real wall. AI Studio stopped us after 10 Argon runs in a day. Plan the run count before you start.
The result in one board
We ran Argon on the same ten tasks as our Hydrogen class, at the levels Argon offers, and graded each answer by running it. The board below is the whole result; the rest of the class is what it means and what it costs.
Gemini 4 Hydrogen
4 thinking levels, Low to Extra High. $1.50 in and $7.50 out per million tokens. Free tier on the API.
Gemini 4 Argon
3 levels in AI Studio, Low to High, opening on Medium. $2.00 in and $10.00 out until November 30. No free API tier.
ReportedList price, not spendread Nov 15, 2026 Levels from the AI Studio run settings; prices from Google’s pricing page, copied the same day.
Does the board say this?
Each line is a conclusion someone might draw from the scoreboard. Pick PASS if the board supports it, FAIL if it does not. A wrong pick shows why.
Argon passed more tasks than Hydrogen, so it is the stronger coder.
On these tasks, Argon is not worse than Hydrogen.
Hydrogen fails the CSV task at High.
The two models disagree on the same edge case.
Argon at High is the safe default for small coding tasks.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 5 readings of one board Each answer follows from the scoreboard in module two.
How it was tested, at no cost
No API key and no bill. Hydrogen’s numbers come from our previous class: AI Studio’s web console, a fresh chat per task, the level set and read back, grounding off. Argon ran two ways. In AI Studio, the same way, until the free daily quota stopped it after 10 runs at Low. Then in Antigravity, Google’s agent app, which has its own usage limit: one conversation per level, the ten prompts in a file, each answer written to its own file with nothing run.
Every answer, from either model and either app, went through the same grader on our machine: run against fixed cases, after a stub for each task had failed. One task’s checks were rebuilt for this class so that each check fails on its own broken answer; two of the old ones could never fail.
What Antigravity cannot give you
A clean time per task. At Low it thought once and wrote the ten files a median 1.4 seconds apart; at Medium the gap was 22.7 seconds and at High 41.3. Those gaps include the agent’s own steps, so they bound the work rather than time it. The per-task times on this page are from AI Studio only.
Measured How the runs were made
- Measured
- Pass or fail per task and level; wall time and tokens where the app shows them.
- Method
- AI Studio web console with no API key, a fresh chat per run, level read back, grounding off; Antigravity, one conversation per level, answers as files, nothing executed by the model. Every answer graded by execution against fixed cases.
- n
- One run per cell; no confidence interval is claimed.
- Source
- The saved runs and answer files, Nov 15, 2026.
- Excludes
- Billed cost: free use bills nothing. Argon Medium and High in AI Studio, which the daily quota stopped.
Is this model comparison fair?
Five things a head-to-head needs before its result means anything, in order. For each setup, pick the first one it lacks; if it has all five, pick the last button.
You paraphrase the prompt for the second model to suit it
- Same tasks, same wordsBoth models get the identical prompts.REFUSED
- Level set on purposeEach run states its thinking level and reads it back.never reached
- Graded by running the answerCases you wrote, executed, not a model judging a model.never reached
- A grader that can failEach check shown failing on a broken answer.never reached
- Surface recordedWhich app ran it, since the same model varies between apps.never reached
- Trust the comparisonUse the result.never reached
Same prompts; you leave each app on its default level
- Same tasks, same wordsBoth models get the identical prompts.pass
- Level set on purposeEach run states its thinking level and reads it back.REFUSED
- Graded by running the answerCases you wrote, executed, not a model judging a model.never reached
- A grader that can failEach check shown failing on a broken answer.never reached
- Surface recordedWhich app ran it, since the same model varies between apps.never reached
- Trust the comparisonUse the result.never reached
Levels set; a third model rates which answer is better
- Same tasks, same wordsBoth models get the identical prompts.pass
- Level set on purposeEach run states its thinking level and reads it back.pass
- Graded by running the answerCases you wrote, executed, not a model judging a model.REFUSED
- A grader that can failEach check shown failing on a broken answer.never reached
- Surface recordedWhich app ran it, since the same model varies between apps.never reached
- Trust the comparisonUse the result.never reached
Executed against cases nobody checked can fail
- Same tasks, same wordsBoth models get the identical prompts.pass
- Level set on purposeEach run states its thinking level and reads it back.pass
- Graded by running the answerCases you wrote, executed, not a model judging a model.pass
- A grader that can failEach check shown failing on a broken answer.REFUSED
- Surface recordedWhich app ran it, since the same model varies between apps.never reached
- Trust the comparisonUse the result.never reached
All four, but you mix two apps without saying which ran what
- Same tasks, same wordsBoth models get the identical prompts.pass
- Level set on purposeEach run states its thinking level and reads it back.pass
- Graded by running the answerCases you wrote, executed, not a model judging a model.pass
- A grader that can failEach check shown failing on a broken answer.pass
- Surface recordedWhich app ran it, since the same model varies between apps.REFUSED
- Trust the comparisonUse the result.never reached
Same prompts, levels read back, executed, checks proven, surface on every cell
- Same tasks, same wordsBoth models get the identical prompts.pass
- Level set on purposeEach run states its thinking level and reads it back.pass
- Graded by running the answerCases you wrote, executed, not a model judging a model.pass
- A grader that can failEach check shown failing on a broken answer.pass
- Surface recordedWhich app ran it, since the same model varies between apps.pass
- Trust the comparisonUse the result.REFUSED
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 5 conditions, 6 setups The conditions this class used, and the failures it met along the way.
The scoreboard: both models, every level
Each cell is one answer, executed against the task’s cases; the figure is checks passed out of checks run. Hydrogen’s columns are AI Studio; Argon’s first three are Antigravity, and the last is Argon at Low in AI Studio, for the same model in the other app.
| Task | Hydrogen Low | Medium | High | Argon Low | Medium | High | Argon Low, AI Studio |
|---|---|---|---|---|---|---|---|
| Glob match | 8/8 | 8/8 | 8/8 | 8/8 | 8/8 | 8/8 | 8/8 |
| CSV record | 6/6 | 6/6 | 5/6 | 6/6 | 6/6 | 6/6 | 5/6 |
| Interval minus | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
| Banker’s round | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
| Topo sort | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
| Palindrome | 7/7 | 7/7 | 7/7 | 7/7 | 7/7 | 7/7 | 7/7 |
| Strip comments | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
| Version compare | 6/6 | no clean run | 6/6 | 6/6 | 6/6 | 6/6 | 6/6 |
| Hard wrap | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 | 5/5 |
| LRU cache | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 | 4/4 |
Measuredn = 1 run per cell Green passed every check; red missed at least one. Hydrogen from Class 110’s first pass; Argon from Nov 15, 2026. Version compare at Medium has no clean Hydrogen run. Topo sort uses the rebuilt checks for both models; Hydrogen’s answers were re-run against them.
Where the board is red, it is the same task: the CSV record, where an empty line should give one empty field. Hydrogen at High returned no fields at all. Argon at Low in AI Studio missed the same check. Nothing else separated the two models on these tasks.
Read it for what it is: one run per cell, on tasks a strong model should pass. It says Argon is not worse here and not better; it cannot say how often either slips.
Hydrogen’s own class has the full four-level run, Extra High included: Gemini 4 Hydrogen, audited.
Hydrogen against Argon, cell by cell
Each row puts Hydrogen’s count beside Argon’s for one level. HOLDS means Argon matched or beat Hydrogen at that level; FAILS means it fell short.
Low
measured delta 0 of 10 0.0%
Medium
measured delta +1 cells +11.1%
High
measured delta +1 of 10 +11.1%
Low, AI Studio only
measured delta -1 of 10 -10.0%
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 4 level pairs Counts from the scoreboard; one run per cell.
Same model, two apps, two answers
Argon at Low ran the ten tasks twice: once in AI Studio, a fresh chat per task, and once in Antigravity, all ten in one conversation. AI Studio’s answers passed 9 of 10; Antigravity’s passed 10 of 10. The difference is one task, the CSV record, on the empty-line case.


Measuredn = 1 saved chat Screenshots of the saved AI Studio chat, Nov 15, 2026; the tooltip is AI Studio’s own.
Two things change between the apps, and one run cannot pull them apart: the wrapper each app puts around the prompt, and plain sampling, since the same model can answer the same prompt two ways. Either way the lesson is practical. Grade the answer you got, in the app you will use, and do not carry a result from one surface to the other.
Find the cell that changed with the app
Same model, same level, same prompts, two apps. One task came out differently. Find it.
Nine tasks agree in both apps
Which task passed in one app and missed in the other?
Same app, two models
Which task separated them?
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 6 answers From the graded runs in both apps.
Speed and tokens at Low
Only AI Studio shows time and tokens per run, and only Low ran there for both models. At Low, Argon took a median 7.6 seconds a task against Hydrogen’s 5.3, and used 0.38× Hydrogen’s output tokens over the tasks both have counts for. Thinking is billed as output on both models.
| Task | Hydrogen, s | Argon, s | Hydrogen, tokens | Argon, tokens |
|---|---|---|---|---|
| Glob match | 5.6 s | 8.6 s | not read | 267 |
| CSV record | 9.1 s | 33 s | 1021 | 331 |
| Interval minus | 7.1 s | 10.3 s | 1111 | 336 |
| Banker’s round | 4 s | 6.1 s | 511 | 214 |
| Topo sort | 4.4 s | 5.5 s | 711 | 310 |
| Palindrome | 4.4 s | 7.9 s | 825 | 221 |
| Strip comments | 6.4 s | 7.3 s | 921 | 390 |
| Version compare | 5 s | 8.3 s | 1000 | 274 |
| Hard wrap | 5.6 s | 5.7 s | 1064 | 303 |
| LRU cache | 3.6 s | 7.1 s | 506 | 528 |
Measuredn = 1 run per cell Wall time and output tokens (thinking included) from AI Studio’s run records and token tooltip. Glob match has no Hydrogen token count: its chat link was not recorded.
The CSV record is the outlier: Argon took 33 seconds on it, against 9.1 for Hydrogen, and still missed the empty-line case. Everything else sat within a few seconds. At this size, speed is not a reason to pick either model.
Build the comparison prompt
Start from the vague ask and switch clauses on. Each clause says what it closes.
Compare these two models on some coding tasks.
Paraphrasing changes the test.
Defaults differ between models and apps.
A model that tests its own code is grading itself.
The same model varied between apps here.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 4 clauses Each clause is a rule from the RUN file this class used in Antigravity.
What Argon costs, per token and per task
Google’s pricing page lists Argon at $2.00 per million input tokens and $10.00 per million output tokens through November 30, 2026, then $4.00 and $20.00. Hydrogen is $1.50 and $7.50, with no date on it. So Argon is 1.33× Hydrogen per token now, and 2.67× from December. The API page lists no free tier for Argon; the AI Studio console is free with no key selected.
| Model, level | Index | Cost a task | Minutes a task |
|---|---|---|---|
| Hydrogen, Medium | 41 | $0.80 | 3.8 |
| Hydrogen, High | 51 | $1.87 | 8.3 |
| Argon, Medium | 48 | $4.31 | 30.0 |
| Argon, High | 50 | $4.97 | 29.6 |
| Carbon, High | 60 | $5.48 | 31.3 |
Reportedn = 5 rows, read Nov 15, 2026 From the Artificial Analysis Google table, copied verbatim the same day. Their figures move; date every one you quote.
Per task the gap widens. On that table Argon at High costs 2.66× Hydrogen High per task and takes 3.57× as long, for an index score one point lower. A benchmark task is not your task, but the direction matches our board: on work Hydrogen can do, Argon buys nothing you can see.
Read Argon’s price row
Lines copied from Google’s pricing page on the day. Open each to see what it says, and what it does not.
Feature Free Tier Paid Tier, per 1M tokens in USDYour answers stay in this browser. Nothing is sent anywhere. Reset clears them.
List price, not spendn = 1 pricing page, read Nov 15, 2026 Copied from the page’s table text; spacing tidied.
What a month of tasks costs, by model
Pick a model and level and slide the number of tasks. The cost per task is the benchmark table’s, so read it as a benchmark-sized task, not yours.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
List price, not spendReportedn = 5 rows, read Nov 15, 2026 Cost per task from the benchmark table; the tool multiplies by your count.
Carbon, from the page only
Gemini 4 Carbon was locked on both free surfaces: AI Studio shows it behind a paid-access prompt, and Antigravity lists it as unavailable. So everything here is read, not tested. Its list price matches Argon’s exactly, $2.00 and $10.00 until November 30. On the benchmark table it scores 60 at High, the highest of the three, at $5.48 and 31.3 minutes a task. Its model card reports 88.2% on SEC-bench Pro.
Read is not tested
None of the Carbon figures on this page came from a run of ours. They are what Google and Artificial Analysis published, copied on the day. Treat them as claims until you can run the model on your own tasks.
Did we run it, or read it?
Each figure is on this page. Pick PASS if it came from a run of ours, FAIL if it was read from someone else’s page.
Carbon scores 88.2% on SEC-bench Pro.
Argon at Low passed 9 of 10 tasks in AI Studio.
Argon at High costs $4.97 a task.
Argon took a median 7.6 s a task at Low.
Carbon costs $10.00 per million output tokens.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
MeasuredReportedn = 5 figures Each answer is the label the figure carries on this page.
The free limits are the real wall
Testing for free meant testing inside three different limits, and each one decided the shape of this class.
- AI Studio, daily quota. The card “You’ve reached your quota for the day” stopped Hydrogen after 21 completed runs on the day, and Argon after 10. The card offers an upgrade or an API key; we clicked neither.
- Antigravity, weekly limits. Its usage panel showed 77% of the weekly Gemini limit and 26% of a separate weekly Argon limit left before our Medium and High runs.
- Levels. Argon offers 3 levels in AI Studio; Hydrogen offers 4. There is no Extra High column for Argon because there is no Extra High.

Quotedn = 1 screenshot The model menu and usage panel as they read before the Medium and High runs, cropped to the menus.
Count runs before you start: tasks times levels times repeats. Thirty Argon runs did not fit one AI Studio day; in Antigravity they fit in one evening.
MeasuredQuotedread Nov 15, 2026 Quota cards and the usage panel as they read on the day; limits change without notice.
How many runs will your comparison need?
Pick how many levels each model offers and slide the task count; the tool gives runs for one pass of one model. Double it for two models, and again for a second pass.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 1 quota stop per model Thresholds come from where AI Studio stopped us on the day.
Pick your model, and keep measuring
On checkable coding tasks of this size, start with Hydrogen at Low, grade by running the answer, and raise the level only when a check fails. Reach for Argon where its model card points: its card reports 77.9% on DeepSWE and 84.2% on a long-context retrieval test up to a million tokens, work our tasks never touch. Price that work per task before you commit.
Which model for this job?
Answer three questions; the list keeps only the advice that fits what was measured here.
- Start with Hydrogen at LowOn checkable tasks like these, both models passed; the cheaper one at its shallowest level is the place to start.
- Trial Argon on your own long tasksThat is where its model card claims an edge; our tasks never tested it.
- Price it per task before you commitPer task, Argon High cost 2.66× Hydrogen High on the table.
- Grade by running the answerEvery answer here was executed against cases; that is what found the misses.
- Note the app on every resultThe same model answered differently in two apps.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
MeasuredReportedn = 10 tasks, 3 levels Advice from this class’s board and the tables it read; not your workload.
Generate your model comparison file
Fill in the two models and your task set; the file lists the steps this class ran, in order.
Your answers stay in this browser. Nothing is sent anywhere. Reset clears them.
Measuredn = 8 steps The steps this class ran, in its own order.
Questions
Is Gemini 4 Argon better than Gemini 4 Hydrogen?
Not on our 10 small coding tasks: both passed nearly every one at every level we ran. On the benchmark table we read, Argon at High scores 50 against Hydrogen High's 51, at $4.97 a task against $1.87.
How much does Gemini 4 Argon cost?
$2.00 per million input tokens and $10.00 per million output tokens through November 30, 2026, then $4.00 and $20.00 from December 1. Hydrogen is $1.50 and $7.50.
What thinking levels does Gemini 4 Argon have?
In AI Studio, 3: Low, Medium and High, with Medium selected on a new chat. Hydrogen has 4, including Extra High.
Can I use Gemini 4 Argon for free?
In the AI Studio web console with no API key selected, and in Antigravity, yes, within limits. The API pricing page lists no free tier for Argon. AI Studio stopped us after 10 Argon runs on the day.
Can I test Gemini 4 Carbon?
Not for free on our accounts: AI Studio shows it behind a paid-access prompt, and Antigravity lists it as unavailable. This class covers Carbon from its model card, the price page and a benchmark table only.
Why did the same model give different answers in AI Studio and Antigravity?
Each run samples a fresh answer, and the two apps wrap the prompt differently. Antigravity also wrote all ten answers in one conversation. One run per cell cannot tell you a rate, only that it varies.
How did you grade the answers?
Every answer was executed on our machine against fixed cases, after a stub for each task was shown to fail. One task's checks were rebuilt so each check fails on its own broken answer.
Which should I use, Argon or Hydrogen?
For small, checkable coding tasks like these, start with Hydrogen at Low and measure. Reach for Argon on the work its model card targets, long agentic jobs and very long context, and price it per task, not per token.