This article checks public sources as of November 16, 2026, focusing on Hydrogen, released November 12, and Carbon's wider availability on November 10. We read official model cards and methods, evaluation providers' original pages and community posts. We did not call paid models or rerun those leaderboards. The prompts and offline records are original test designs, not measured model results.
Google: Introducing Gemini 4 Hydrogen · Google: Gemini 4 Carbon is now available
Place Hydrogen, Carbon and Argon in their respective roles
This release suggests a division of work: Hydrogen emphasizes efficient everyday agent execution, Carbon addresses harder reasoning over longer workflows, and Argon remains in established workflows. A model's announcement, restricted access and wider release are separate events. Align their dates before comparing older experiences.
Google: Introducing Gemini 4 Hydrogen · Google: Gemini 4 Carbon is now available
| Model | Date or role | Check before use |
|---|---|---|
| Gemini 4 Hydrogen | Released 2026-11-12; everyday execution | Default high; low, medium and xhigh are also available |
| Gemini 4 Carbon | Wider availability 2026-11-10; earlier restricted access | Harder tasks, longer verification loops and larger resource budgets |
| Gemini 4 Argon | An existing Gemini 4 model | Google still identifies it as suitable for most engineering and knowledge work |
Google: Introducing Gemini 4 Hydrogen · Google: Gemini 4 Carbon is now available
Our selection framework starts with the task, then chooses a model and level. Formatting, bounded fixes and extraction can establish a faster configuration's baseline. Architecture decisions, constraints across files and elusive bugs justify comparing deeper reasoning. Judge whether the deliverable meets its requirements, rather than the length of the answer.
A model name does not guarantee the same abilities in every product
| Entry point | Hydrogen | Carbon |
|---|---|---|
| Gemini app | Across tiers; free users gain access through appropriate automatic selection | The release lists Google AI Ultra |
| AI Studio / Gemini API | Free and paid access have their own quotas | Paid developer access |
| Google Antigravity | The release lists paid AI Pro / Ultra | The release lists AI Ultra |
| Search and business products | Check Search, business products and organizational permissions separately | Check Ultra, business entry points and special programs separately |
Google: Introducing Gemini 4 Hydrogen · Google: Gemini 4 Carbon is now available · Google: Hydrogen developer announcement · Gemini Apps release notes
Subscriptions, API billing, IDE agents and organizational permissions have separate boundaries. Selecting a model in chat does not establish an API balance, identical tools or identical context limits. Before buying, record the product, level and delivery type, and match the seller's description to the rights you need. Whether our products include Gemini depends on the current cards and delivery details.
Carbon's release lists introductory public API rates of $2 per million input tokens and $10 per million output tokens, with the introductory period ending November 30. Hydrogen's announcement lists $1.50 input and $7.50 output per million. These are specified API rates, not a conversion into subscription messages or a third-party IDE's quota deductions.
Google: Introducing Gemini 4 Hydrogen · Google: Gemini 4 Carbon is now available
For multimodality and long context, inspect the input-output contract
Both model cards list text, image, audio and video inputs, with text output. Producing a spreadsheet or presentation through tools involves execution, file writing and subsequent inspection. That differs from a model directly emitting audio or video. An attractive application demo also reflects the code and execution environment, which should be assessed separately.
| Item | Condition in the card or methods | Practical verification |
|---|---|---|
| Context | Up to 1M tokens for input and output combined | Budget for source material, tool history and the final response |
| Long output | Hydrogen: up to 64K default synchronous output; up to 1M through background long decode | Check whether your entry point supports that mode |
| Complex files | Visual reading, formulas and final artifacts are distinct stages | Check chart coordinates, cell formulas, page references and exported files |
A larger window admits more material without guaranteeing accurate use of every detail. Give files identifiers, request a file and page for important claims, and ask what chart, unit or condition across pages could overturn the conclusion. Open the generated file instead of relying on a completion statement.
How to test four thinking levels without merely waiting longer
Hydrogen explicitly offers low, medium, high and xhigh, with high as the default. Those options describe this model's configuration; other models, products and automatic routes may use different meanings. When raising the level, record quality, elapsed time and output volume together so that a longer explanation is not mistaken for more completed work.
Google: Introducing Gemini 4 Hydrogen · Google: Hydrogen developer announcement
| Level | Tasks to try first | Upgrade or stop signal |
|---|---|---|
| low | Extraction or formatting with clear output rules | Preserve the failed input and upgrade if constraints are missed |
| medium | Fixes and analysis with a few dependencies | Compare additional correct results with the extra waiting cost |
| high | Multiple steps requiring tool feedback | Verify that tests actually ran and artifacts were read back |
| xhigh | Hard problems with a clear verification method | Set a human checkpoint if further reasoning adds no evidence |
This order is our experimental proposal, not an official guarantee. Retrieval quality, tool permissions, environment failures and contaminated context can also constrain a task. A model waiting on an unavailable tool may gain nothing from a higher level. Inspect the last tool status and error before retrying, changing models or repairing the environment.
Read official scores alongside the harness, version and attempt count
Hydrogen's methods specify pass@1 and the highest thinking setting except where noted. The Terminal-Bench 4.0 and 4.1 details instead specify high, a mini-swe-agent vision harness and average reward across five attempts per task. That average is not the chance of succeeding at least once in five tries. The exceptions also prevent labeling the entire table xhigh.
| Source or evaluation | Configuration to retain | Misreading avoided |
|---|---|---|
| Hydrogen Terminal-Bench | Keep 4.0 and 4.1 separate; high; average reward over five attempts | Mixing versions or treating the mean as pass@5 |
| Hydrogen SWE-sweep | Antigravity CLI 1.4.8; subtask-weighted reward | Treating different agent harnesses as the same bare-model test |
| Carbon comparisons | Some external-model numbers come from provider reports | Calling every column a common independent rerun |
| Carbon AA / Vals | Release methods list AA v5.0 and Vals v2.2 | Drawing a trend line straight through later index revisions |
Hydrogen Model Evaluation methodology PDF · Carbon Model Evaluation methodology PDF
Google's claims about Hydrogen's speed, fewer fabricated claims and collaboration can become hypotheses for your own trial. Fix inputs and task limits, establish equal repair quality, then compare wall-clock time. Only then does a speed gain describe equivalent delivery rather than stopping early or skipping checks.
Artificial Analysis: read capability, time and cost together
| Configuration | Index score | Task statistics on the same page |
|---|---|---|
| Hydrogen Xhigh | 54 | About 14.9 minutes per task; $3.16 average cost |
| Carbon High | 60 | About 33.0 minutes per task; $5.48 average cost |
Artificial Analysis: Hydrogen model page · Artificial Analysis: Carbon model page
This index aggregates nine evaluations. The score belongs to its configuration; the page's average time and cost describe that evaluation workload, not a quote for every chat. Model pages change, so we retain the access date and index version instead of inserting current values into older launch charts.
Artificial Analysis: Hydrogen model page · Artificial Analysis: Carbon model page
A conditional trial strategy follows: expand a higher configuration only when the extra failures it resolves matter to you. For frequent, simple outputs, waiting time may affect usability more than token price. For difficult work, one cheap failure followed by expensive rework may cost more than starting with a stronger configuration.
CursorBench provides workflow evidence, not points to add to AA
| Configuration | Score | Average cost |
|---|---|---|
| Carbon Medium | 60.1% | $8.49 |
| Hydrogen High | 54.3% | $5.59 |
| Argon Medium | 52.8% | $4.39 |
This leaderboard is closer to the software work Cursor defines, while its environment and task distribution still belong to that evaluation. Carbon uses Medium here and High on AA's model page. Different scales, workloads and harnesses mean 60.1% and an AA score of 60 do not validate each other as the same measurement.
Artificial Analysis: Hydrogen model page · Artificial Analysis: Carbon model page · Current Cursor Insights model leaderboard
Leading positions and lower-cost claims in Google's release also belong to their original comparison set. Adding models or updating prices can change the relationship. For procurement or migration, preserve the current leaderboard, thinking level and failures from your own tasks so that a launch-time ratio does not become a permanent promise.
Google: Gemini 4 Carbon is now available · Current Cursor Insights model leaderboard
What a multimodal knowledge-work evaluation measures
MKA is a collaboration between Google DeepMind and Artificial Analysis that asks agents to turn complex visual materials into actual files. AA describes 200 private tasks, a common Stirrup harness and up to 250 turns. Ten public tasks let readers inspect the format; they do not reproduce the complete private leaderboard.
Artificial Analysis: MKA methods and results · Artificial Analysis: MKA evaluation page · Google's public MKA dataset
Its headline Climb WSR measures weighted satisfaction of harder rubric criteria, not the probability of completing a task. Scoring distinguishes baseline Regression checks from harder Climb criteria and uses an agentic judge that opens deliverables. This is closer to artifact quality than final text alone, but still depends on rubrics, automated review and domain distribution. The collaboration should remain disclosed.
Artificial Analysis: MKA methods and results · Artificial Analysis: MKA evaluation page
The useful transfer is to replace 'the file was generated' with checks that it opens, formulas and sources are correct, conclusions match the original figures and edits stay in scope. Make acceptance a hard gate, then use component scores to locate failures. We did not run MKA.
Why community reports contain both surprises and stalls
| Original post | Reported experience | What it cannot establish |
|---|---|---|
| Early Carbon development-pipeline trial | Four positive phases; the author explicitly notes the tiny sample | Effort follows difficulty, preventing a causal medium/high comparison |
| A Hydrogen user's extended workflow | Useful engineering work reported; another model's review found problems | One project and model-assigned grades do not establish general reliability |
| Cursor Carbon High issue report | Repeated continue prompts followed by a Thinking stall; version details supplied | The report alone cannot isolate the model, IDE, network or server cause |
Original community post: early Carbon pipeline results · Original community post: Hydrogen experience and review · Cursor community: Carbon High stall report
These reports help construct test cases and troubleshooting steps; they do not measure the proportion of all users affected. Positive workflows often already contain contracts, tools and review. Negative reports may combine model and product failures. Ask whether inputs match, which level was used, whether tool evidence was retained and whether artifacts were actually inspected.
When work stalls, retain the scope, model label, product version and last tool status. Remove credentials and private materials before contacting support. For a claimed completion with missing results, inspect files, tests and diffs directly. Endless continue prompts are a poor substitute for diagnosis; each retry needs a specific change and stopping condition.
A reusable trial: common tasks, fixed budgets and separate checks
Choose three tasks using materials you own or may use: a CSV with anomalous rows, a bounded code fix and a chart report requiring page evidence. Freeze input copies and acceptance criteria. Reset context independently for several runs per configuration, retaining failures and rework. With a small pilot, report raw counts without presenting a percentage as a stable success rate.
Task: [describe one verifiable task]
Inputs: [fixed file list, versions or checksums]
Allowed edits: [explicit files and directories]
Prohibited actions: [unneeded deployment, messaging, payment, etc.]
Acceptance criteria:
1. [functionality and edge cases]
2. [executable test or manual read-back method]
3. Preserve original inputs and files outside the scope.
List missing information first; proceed when sufficient. Record permission or environment blockers and completed work.
Return actual edits, checks genuinely run and their outputs, unresolved items and artifact paths.
Do not report planned checks as completed.Review the existing artifacts without trusting the author's completion claim.
Task contract: [paste frozen objectives and acceptance criteria]
Evidence: [original inputs, artifacts, diffs and test output]
For each criterion return: criterion | source evidence location | check performed | pass/fail/unverified.
Prioritize consequential errors, units and formulas, source support and edits outside scope.
Separate pre-existing evidence from checks you add now.
If files or the execution environment are unavailable, mark the conclusion unverified. Do not grade fluent prose as correctness.Offline code: count retries and failures in the cost of accepted work
A and B below are synthetic candidates, not real models. A run is accepted only when tests, artifact read-back and scope checks all pass. Total cost includes every attempt and is divided by accepted runs. One deliberate failure illustrates the difference between a cheap attempt and a usable result.
import json
from decimal import Decimal
from statistics import median
# Synthetic teaching fixture. A and B are not Google model results.
rows = [
("A1", "A", True, True, True, 10, "0.04"),
("A2", "A", False, True, True, 12, "0.05"),
("A3", "A", True, True, True, 14, "0.06"),
("B1", "B", True, True, True, 20, "0.12"),
("B2", "B", True, True, True, 24, "0.15"),
("B3", "B", True, True, True, 30, "0.18"),
]
assert len({row[0] for row in rows}) == len(rows)
report = {"synthetic": True, "model_calls": 0, "candidates": {}}
for candidate in ("A", "B"):
runs = [row for row in rows if row[1] == candidate]
assert len(runs) == 3
assert all(type(v) is bool for row in runs for v in row[2:5])
assert all(row[5] > 0 and Decimal(row[6]) >= 0 for row in runs)
passed = sum(all(row[2:5]) for row in runs)
spend = sum((Decimal(row[6]) for row in runs), Decimal("0"))
report["candidates"][candidate] = {
"runs": len(runs), "accepted": passed,
"all_attempt_cost": str(spend),
"cost_per_accepted": str((spend / passed).quantize(Decimal("0.001"))) if passed else None,
"median_seconds": median(row[5] for row in runs),
}
assert report["candidates"]["A"]["accepted"] == 2
assert report["candidates"]["A"]["cost_per_accepted"] == "0.075"
assert report["candidates"]["B"]["accepted"] == 3
assert report["candidates"]["B"]["cost_per_accepted"] == "0.150"
print(json.dumps(report, ensure_ascii=False, sort_keys=True))Expected results: A accepts 2 of 3 runs at 0.075 per accepted result; B accepts 3 of 3 at 0.150. Median times are 12 and 24 seconds. Every number comes from the fixture. Real trials should also count human review, recovery and tool costs. When no run succeeds, report the ratio as unavailable instead of dividing by zero or hiding failures.
Turn your test results into a practical choice
| Task characteristics | Compare first | Condition for wider use |
|---|---|---|
| Frequent tasks with clear rules | Different Hydrogen thinking levels | Accuracy and scope gates pass while total elapsed time falls |
| A few difficult tasks with long dependencies | Carbon and established strong models under a fixed budget | Extra failures resolved justify additional time and cost |
| A stable existing Argon workflow | Keep its baseline and test new models on its failure cases | Gains reproduce across independent runs and migration cost is acceptable |
A multi-model workflow can use a stronger model for decomposition, an execution model for bounded tasks, then executable checks and separate review to merge results. Clear interfaces and acceptance conditions are prerequisites. Context transfer, repeated retrieval and synthesis errors can make extra agents more expensive. Treat orchestration as a comparison candidate rather than an automatic upgrade.
Before buying, match your required entry point to the product: your own or a separate account, subscription or API, supported products and delivery timing. A model announcement does not establish that a particular item includes every model permission. Our current product pages define what is offered and supported.
Connect model news to a repeatable acceptance loop
Keep the old baseline, inputs, versions and failures for every upgrade. Reuse the same materials at the next model or product change to identify which stage improved.
- Confirm the model, product, level and date
- Freeze inputs, budget and acceptance criteria
- Execute while retaining tool evidence
- Open artifacts, run checks and inspect scope
- Count every attempt and its cost; continue or roll back
Evidence to prioritize at the next update
Track improvements under the same harness and budget, alongside product fixes for quotas, routing and stalls. A higher benchmark score warrants another trial; a recurring failure finally being resolved may determine migration. A small fixed acceptance set, source versions and recovery costs turn updates into explainable decisions.
- Did the index version or model snapshot change?
- Are the harness, thinking level and budget comparable to the previous run?
- Do the new results reproduce on your own failure cases?
- Is there official confirmation and an inspectable fix for a stall or metering issue?
We distinguish vendor claims, independent measurements, individual reports and original experimental designs. Public scores can change. Check current official documentation and product delivery conditions for entry points, charges and rights.