AI DevelopmentDecision Matrix8 min readPublished November 13, 2026

Select the model for the work, then define when to escalate

Gemini 4 Hydrogen vs Argon vs Carbon: Which One to Use

Choose between Gemini 4 Hydrogen, Argon and Carbon by task, thinking level and total cost, with current prices and the December increase separated.

DA
Digital Applied Team
Senior strategists
PublishedNovember 13, 2026
Read time8 min
SourcesGoogle API documentation and Artificial Analysis

Choose Gemini 4 Hydrogen as the first candidate for routine coding, document work and tool execution; compare Argon for sustained difficult work, and Carbon when the hardest reasoning is worth additional time and spend. That starting point follows Google’s model-selection guidance, not a universal ranking. A useful choice still depends on whether the model completes your particular task correctly.

The practical complication is that model names are only part of the configuration. Thinking settings change behavior, public evaluations use particular settings, and Carbon and Argon have a scheduled price increase after November. A fair comparison needs to hold the task and acceptance criteria steady while recording the complete cost of an accepted result. Otherwise, an inexpensive-looking default can conceal retries and review work.

Key takeaways
  1. 01
    Begin with the taskEveryday execution and difficult architectural judgment need different evaluation examples, even when both happen in one project.
  2. 02
    Compare explicit settingsHydrogen defaults to high; Argon and Carbon default to medium. A default-versus-default test is not a matched thinking-level test.
  3. 03
    Forecast both price periodsCarbon and Argon standard rates move from $2/$10 to $4/$20 per million input/output tokens on December 1.
  4. 04
    Keep escalation accountableA second model can help with a hard decision, but handoffs need clear inputs, stopping rules and evidence.

01 — Workload fitUse three candidates for three kinds of work

Google describes Hydrogen as an efficient option for everyday coding, interactive applications, tool-using agents and document workflows. It positions Carbon as the strongest model for difficult reasoning, and Argon as an option for advanced long-horizon tasks with lower token use than Carbon. Those descriptions are useful hypotheses to test; they are not guarantees about a particular codebase or document collection.

A hypothetical product team may need to update a straightforward form, diagnose an intermittent data-loss bug and decide how to restructure a shared service. The first task has a relatively clear output and acceptance check. The other two involve uncertainty about causes or trade-offs. Routing all three by the number of words in the request would miss the distinction that actually matters.

Use the Hydrogen launch analysis for its detailed limits and cost model. For this comparison, keep the first decision narrower: which candidate deserves a trial for each class of work? That approach allows a team to select more than one model without building an elaborate routing system before it has useful evidence.

Hydrogen
Everyday execution
Clear tasks and useful checks

Start here for bounded edits, source-based document work and tool sequences whose results can be verified directly.

Candidate default
Argon
Sustained difficult work
Longer reasoning with a budget

Compare it when a task needs several connected decisions and a lower token footprint than Carbon would be valuable.

Evaluate the trade-off
Carbon
The hardest judgments
Complex causes and edge cases

Test whether deeper reasoning materially improves the result enough to justify its time, output and review costs.

Escalation candidate

02 — Price periodsMake December visible in the comparison

The Gemini API rate card gives Hydrogen standard rates of $1.50 input and $7.50 output per million tokens. Argon and Carbon share introductory rates of $2 and $10 through November 30, rising to $4 and $20 on December 1. Output billing includes thinking, which matters when a model spends more tokens reasoning before its visible answer.

A November prototype can therefore make the more expensive options look closer to Hydrogen than they will be in a December operating budget. Keep the introductory and ongoing periods in separate columns. The forecast should use the period when the workload will run, rather than whichever number made the initial trial easiest to approve.

The Carbon pricing guide explains the public-access transition in more detail. For a selection decision, the important point is that equal token rates for Argon and Carbon do not imply equal task cost. Their token consumption can differ, and the cheaper result is the one that reaches the agreed acceptance condition with less total spend.

Standard Gemini API USD rates per million tokens, retrieved November 17, 2026. Hydrogen’s displayed rate has no December change listed here; tools, caching and other consumption options are separate.
ModelThrough November 30From December 1
Hydrogen$1.50 input / $7.50 output$1.50 input / $7.50 output
Argon$2 input / $10 output$4 input / $20 output
Carbon$2 input / $10 output$4 input / $20 output

03 — Thinking levelsAvoid an accidental comparison of defaults

Hydrogen offers low, medium, high and xhigh thinking, with high as the default. The Argon guide and Carbon guide list low, medium and high, defaulting to medium. Record that setting explicitly. A test that changes both model and thinking depth can answer a product-choice question, but it cannot isolate the effect of the model alone.

There are two useful comparisons. One asks which configuration gives the best practical outcome for a workload, allowing each model its best tested setting. The other holds comparable settings steady to investigate a specific behavior. Label which comparison you are doing. Mixing them can turn a valid observation into an unjustified claim that one model is intrinsically faster or more economical.

For a simple extraction task, additional thinking may add cost without changing the accepted result. For an ambiguous bug, it may prevent a mistaken fix. Test both cases rather than selecting a setting from its name. Also preserve output limits and timeouts: a candidate cut off by an inherited limit has not received the same opportunity to complete the task.

Record the complete configuration

Store the model ID, thinking level, tools, relevant limits and acceptance criteria with each run. A result labelled only “Gemini 4” is too ambiguous to reproduce.

04 — Independent evidenceUse the index to shortlist, not to certify

Artificial Analysis reports different index results for these models at its displayed settings. The table below uses the Hydrogen, Argon and Carbon pages from the same retrieval date. Hydrogen is shown at xhigh, while the other two are shown at high. This is a comparison of those evaluated configurations.

The index score is not a percentage accuracy figure and does not forecast the success rate of your own application. Its reported cost per task belongs to the evaluator’s task mix. That makes it useful context for choosing what to test, but a poor substitute for a business-specific acceptance set. A model can lead an aggregate measure and still make an error your workflow cannot tolerate.

Avoid combining a current score with an older score copied from a previous article. Evaluations and displayed model configurations can change. Keep the retrieval date and evaluator attached to every copied number, and retain the settings in a slide or procurement table. Removing those qualifiers makes the comparison look simpler by discarding the information needed to interpret it.

Artificial Analysis model pages, retrieved November 17, 2026. These are evaluator results, not Digital Applied tests or fixed prices for arbitrary work; the settings are not identical.
Evaluated configurationIntelligence IndexCost per index task
Hydrogen xhigh54$3.16
Argon high50$4.97
Carbon high60$5.48

05 — Task economicsCount accepted results instead of cheap calls

An illustrative call using 30,000 uncached input tokens and 6,000 billed output tokens costs $0.09 on Hydrogen at standard rates. With the same assumed usage, Argon or Carbon costs $0.12 during the introductory period and $0.24 from December. The calculation is deliberately simple: it holds token usage constant so the price difference is visible. Real model runs will not necessarily use the same number of tokens.

Now add the possibility of a failed attempt. Two $0.09 calls cost $0.18 before tool and infrastructure charges. That may still be worthwhile if the result is accepted; it may be poor value if the task then needs substantial manual repair. Conversely, a more expensive first attempt can be economical if it reliably avoids that repair. Measure the observed path rather than declaring the lower rate the lower cost.

Record the number of accepted tasks, total model spend and time spent reviewing or correcting results. Include failed attempts in total spend; divide that spend by accepted tasks. For success rate, include every attempt in the denominator. A comparison that removes the difficult runs can make every model look production-ready. If quality differs materially across task types, report separate categories instead of averaging away the difference that should guide routing.

Illustrative calculation

Hydrogen: 30,000 ÷ 1M × $1.50 + 6,000 ÷ 1M × $7.50 = $0.09. Argon/Carbon: $0.12 through November, $0.24 from December. These are assumed token quantities, not measured model usage.

06 — Model handoffsEscalate a decision without duplicating the whole job

A mixed workflow can use one model for an uncertain decision and another for bounded execution. Google suggests pairing Carbon with Hydrogen in this way. The useful design question is where the handoff happens and what evidence moves with it. If the second model must rediscover everything, the extra coordination can erase the expected benefit.

For a hypothetical software change, the planning step could identify the affected interfaces, the unresolved risks and the tests that must pass. The execution step receives that brief and returns a diff, test results and any blocked assumption. Escalation should occur when a stated condition is met, such as a failed premise or an unexpected dependency, rather than whenever an answer sounds less confident.

A second model is also not automatically an independent verifier. If it receives only the first model’s explanation, it may inherit the same missing premise. Give a reviewer access to the relevant source and an explicit acceptance question. Keep the authority to make a consequential change separate from the ability to suggest it. The handoff should make responsibility clearer, not distribute it until nobody owns the result.

Useful handoff contents

Pass the goal, source evidence, decisions already made, unresolved assumptions and acceptance checks. A long transcript without those distinctions is a poor substitute for a clear brief.

07 — Integration limitsCheck the route as carefully as the model name

The models share broad family capabilities, but the application still depends on a particular API, tool configuration and access arrangement. The Hydrogen access guide separates consumer app plans, developer APIs and enterprise routes. A model available in an app is not evidence that the production project has the required API entitlement or quota.

For a document workflow, test the input types and output format that the application actually uses. For a long-running agent, test interruption and recovery. For a tool-driven application, test a missing result and a rejected action. These cases can expose an integration assumption that a generic reasoning benchmark never exercises.

Keep a stable model identifier and a versioned configuration in the evaluation record. If a route or alias changes, rerun the relevant acceptance set before interpreting the result as continuity. The Argon launch guide provides further model context, but current implementation details should come from the documentation for the exact route being deployed.

A route-specific check

Confirm the actual request, returned model information, usage fields and tool behavior in the intended project. Do not infer production access from a successful conversation in a different product.

08 — Selection policyKeep a small policy that the team can explain

Begin with a simple rule: one candidate default for the common workload, one escalation route for the difficult class and a person responsible for exceptions. Expand the policy only when observed failures justify another branch. A complicated router built before evaluation creates more places for assumptions to hide.

A useful decision record states which configuration was selected, which tasks it passed, which failures remain and which price period the forecast uses. Give the team a reason to revisit the choice, such as a material cost change or a recurring failure in an important task category. Avoid replacing a model solely because a new index ranking appears.

AI transformation services can connect model selection to the workflow’s practical outcome. The goal is a repeatable process that completes useful work with acceptable quality, cost and oversight. A model comparison becomes valuable when it changes that process in a way the team can measure and support.

A decision worth keeping

“Hydrogen handles these checked tasks; this failure condition escalates to Carbon; this person owns exceptions” is a usable policy. “We use whichever model looks best today” is difficult to operate.

Your next step

Choose a default and write the escalation rule

Hydrogen, Argon and Carbon are candidates for different combinations of reasoning, speed and cost. Compare explicit configurations on the tasks that matter, include the December price period and measure accepted results. Start with a small selection policy whose exceptions are understandable, then change it when new evidence justifies the work.

Digital Applied

Choose models around measurable work

Design a practical evaluation and routing policy that connects model quality, total cost and the outcome your team needs.

Representative evaluationsClear cost assumptionsAccountable rollout
Start with one decision

Start with a useful comparison

  • →Define the accepted result
  • →Test the relevant configurations
  • →Record the escalation rule
Common questions

Questions, answered.

Hydrogen is a sensible first candidate for the everyday execution tasks Google targets, but the final choice should come from representative acceptance tests. Keep a separate route for tasks where another model demonstrably improves the outcome.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source