Gemini 4 Carbon and Argon Prices Double on December 1: Recalculate and Cut Your Bill

Google's Gemini 4 Carbon and Argon token rates go from $2/$10 to $4/$20 on December 1, 2026. See what changes, a monthly bill formula, and the trade-offs of Neon, Hydrogen, thinking level, caching and Batch.

A pale blue pigeonhole sorting cabinet with one orange envelope sliding out toward a separate tray on the floor, suggesting part of the traffic being rerouted.

On December 1, 2026, every published Gemini Developer API token rate for Gemini 4 Carbon and Gemini 4 Argon doubles. Standard goes from $2 to $4 per million input tokens and from $10 to $20 per million output tokens (thinking included); cached input goes from $0.10 to $0.20. Batch, Flex and Priority double too. Neon ($3/$15) and Hydrogen ($1.50/$7.50) carry no date change. If you keep the same model, tier and token mix, your December token bill for these two models is exactly twice your November one. Search grounding fees are not part of that doubling.

This guide covers the change itself and the decision you need to make before it lands. For the base rate tables and per-request accounting, read our Argon pricing breakdown and the Carbon task-cost calculator; this page links to them rather than repeating them.

Sources and labels used below: rates come from Google’s Gemini API pricing page, captured and re-checked live on November 19, 2026. Bill totals are our own arithmetic on synthetic volumes. The thinking-token figures come from 24 small real requests we ran on November 19; that is a tiny sample on easy prompts, and the section explains its limits. Ofox publishes this article. Gemini 4 Argon is available through Ofox as google/gemini-4-argon; we do not state Ofox prices here, see the Argon model page for those.

What exactly changes on December 1

Every Carbon and Argon cell on Google’s pricing page now carries two dates: one rate “through November 30, 2026” and a second “starting December 1, 2026.” Carbon and Argon have identical numbers in all four paid service tiers.

Service tier (USD per 1M tokens)InputOutput, including thinkingCached input (implicit only)
Standard$2.00 → $4.00$10.00 → $20.00$0.10 → $0.20
Batch$1.00 → $2.00$5.00 → $10.00$0.05 → $0.10
Flex$1.00 → $2.00$5.00 → $10.00$0.05 → $0.10
Priority$3.60 → $7.20$18.00 → $36.00$0.18 → $0.36

Arrows read “through November 30 → starting December 1, 2026.” Source: the Carbon and Argon sections of the pricing page, checked November 19, 2026.

A few details matter when you plan around this table:

  • It is a uniform 2× multiplier. Every token rate in every tier doubles, so the ratio between tiers stays the same. Batch and Flex remain 50% of Standard; Priority remains 80% above Standard.
  • Google calls the Carbon rate introductory. The Carbon guide and the choosing-a-model page both describe $2/$10 as introductory pricing through November 30. Argon’s pricing section uses the same two dates without that word on the pricing page; the effect is the same.
  • Neither model has a free tier. Both pricing sections say “not available on the Free tier,” before and after the change.
  • Thinking level does not change the per-token price. The Carbon guide says its low, medium and high levels come “at the same per-token price.” Thinking changes how many output tokens you generate, not the rate.
  • Argon has a Contributor tab. It lists $1.00 input, $5.00 output and $0.05 cached, with no date qualifier. Google describes it as an invite-only private preview, and its “used to improve our products” row says yes. Unless you have been invited, it is not an option, and the page does not say whether that tab will change.

What does not change

  • Search and Maps grounding. The Standard and Priority tabs for Carbon and Argon list 5,000 free grounding requests per month shared across Gemini 3 and newer models, then $14 per 1,000 (Google Search grounding also has a daily cap of 1,500 requests). No date qualifier is attached. If grounding is a large share of your bill, your total rises by less than 2×.
  • Gemini 4 Neon. $3.00 / $15.00, cached $0.15 on Standard; $1.50 / $7.50 on Batch and Flex; $5.40 / $27.00 on Priority. No date qualifiers, and Google’s choosing-a-model page states that Neon has no introductory price.
  • Gemini 4 Hydrogen. $1.50 / $7.50, cached $0.075 on Standard; $0.75 / $3.75 on Batch and Flex; $2.70 / $13.50 on Priority. Hydrogen also keeps its free tier (rate-limited, and free-tier content may be used to improve Google’s products).
  • Model availability. Google’s deprecations page lists “No shutdown date announced” for both gemini-4-carbon and gemini-4-argon as of November 19. The price change is not a retirement notice.
  • Context and API behavior. All four Gemini 4 models still share a 1,048,576-token unified window, and no request schema change is tied to December 1.

The time zone and billing-period question

Google’s pricing page gives calendar dates only. As of our November 19 check it does not say which time zone defines “November 30” and “December 1,” and it does not say which rate applies to a request that straddles midnight, or to a Batch job submitted on November 30 that finishes on December 1 (Google’s optimization guide gives Batch a target turnaround of up to 24 hours). We found no documented answer, so do not plan on squeezing work into the last hours of November. Treat anything that runs on November 30 or December 1 as possibly billed at the new rate, and confirm actual charges in your Cloud Billing reports after the fact.

If your internal budget month does not match the calendar month (for example, a period running November 15 to December 14), split the usage at the change date and price each side separately. A single blended rate will be wrong in both directions.

Recompute your monthly bill

Use one formula per model, service tier and rate period, then add the rows. This is the same structure as our per-request calculators, scaled to a month.

I = total input tokens for the month (including cached tokens), in millions
C = the cached subset of I that the API reported as cache hits, in millions
O = output tokens including thinking, in millions
Pi, Pc, Po = that tier's input, cached-input and output rates per 1M tokens

monthly_token_bill = (I - C) × Pi + C × Pc + O × Po
monthly_total      = monthly_token_bill + grounding and other tool fees

Rules that prevent most mistakes: C must be a subset of I (never add it on top), O must include thinking tokens exactly once, and grounding is billed per search query, not per chat turn. For the field-by-field mapping from API usage records, see the normalization table in the Carbon task-cost guide.

Here is the formula as a short Python script you can copy and edit. It contains only Google’s published rates, checked November 19, 2026.

RATES = {  # USD per 1M tokens: (input, cached input, output incl. thinking)
    "carbon_argon_standard_nov": (2.00, 0.10, 10.00),
    "carbon_argon_standard_dec": (4.00, 0.20, 20.00),
    "carbon_argon_batch_flex_dec": (2.00, 0.10, 10.00),
    "carbon_argon_priority_dec": (7.20, 0.36, 36.00),
    "neon_standard": (3.00, 0.15, 15.00),
    "hydrogen_standard": (1.50, 0.075, 7.50),
}

def bill(input_m, cached_m, output_m, rate):
    """input_m includes cached_m. All quantities in millions of tokens."""
    assert 0 <= cached_m <= input_m
    pi, pc, po = RATES[rate]
    return (input_m - cached_m) * pi + cached_m * pc + output_m * po

# monthly volumes: (total input M, cached subset M, output incl. thinking M)
for name, v in {"small": (20, 0, 2), "medium": (300, 150, 30),
                "large": (3000, 2000, 250)}.items():
    print(name, {r: round(bill(*v, r), 2) for r in RATES})

Worked examples at three volumes

The three workloads are synthetic. They are chosen to show how caching and the output share move the result, not to represent typical customers.

  • Small: 20M input tokens, no cache hits, 2M output tokens per month. Think of a team prototype or an internal tool with short, varied prompts.
  • Medium: 300M input tokens, of which 150M were cache hits, and 30M output tokens. A production assistant with a stable system prompt and tool definitions.
  • Large: 3B input tokens, of which 2B were cache hits, and 250M output tokens. A heavy agent workload that re-sends long context on every turn.
Monthly token bill (USD)SmallMediumLarge
Carbon/Argon Standard, November rates60.00615.004,700.00
Carbon/Argon Standard, December rates120.001,230.009,400.00
Carbon/Argon Batch or Flex, December rates60.00615.004,700.00
Carbon/Argon Priority, December rates216.002,214.0016,920.00
Neon Standard, same token counts90.00922.507,050.00
Hydrogen Standard, same token counts45.00461.253,525.00

Check the medium December row by hand: (300 − 150) × $4 = $600 for uncached input, 150 × $0.20 = $30 for cached input, 30 × $20 = $600 for output, total $1,230. In the large workload, cached input is two thirds of all input but only $400 of the $9,400 December bill; output alone is $5,000. That is typical of reasoning models: output and thinking, not input, usually drive the increase.

The Neon and Hydrogen rows hold token counts fixed, and that assumption is the weak point of any “just switch models” plan. Different models, and different default thinking levels, generate different numbers of thinking tokens for the same request. The next section shows how large that gap can be.

What our 24 test requests showed about thinking tokens

Small sample, not a benchmark. 24 requests in total: two short, easy prompts, n = 4 per row, one day, one route. The numbers below show that thinking tokens can dominate the bill. They do not tell you how Argon, Carbon or Hydrogen will behave on your traffic, and we draw no general conclusion about any model from them.

To see how much thinking moves the output count, we sent two short prompts to gemini-4-argon and gemini-4-hydrogen through Ofox’s Gemini-native endpoint (/gemini/v1beta/models/{model}:generateContent, documented in Ofox’s Gemini API reference) on November 19, 2026. One prompt was a three-box labeling puzzle; the other asked for a fix to a buggy median function. Each configuration ran twice per prompt (n = 4 per row), with maxOutputTokens 32,768. All 24 answers were correct by our check (right box to sample from; even-length median averaged, input list not mutated).

Model and thinking settingMean input tokensMean visible outputMean thinking tokensMean billable outputMedian latency
Argon, no thinkingLevel set (default)9713501356.9 s
Argon low9712601266.9 s
Argon medium9712701277.7 s
Argon high971379851,12121.8 s
Hydrogen, no thinkingLevel set (default high)971276487756.8 s
Hydrogen medium971336087418.0 s

The Argon high row is roughly 8.8× the billable output of Argon medium (1,121 vs 127 tokens), which we round to “about 9×” below. “0 thinking tokens” means the response’s usageMetadata had no thoughtsTokenCount field and totalTokenCount equaled prompt plus visible output. Raw requests and responses are saved with the article’s working files.

At Google’s Standard rates, those means work out to roughly $0.0029 per request for Argon medium at December rates, $0.0228 for Argon high at December rates, and $0.0060 for Hydrogen at its default. That is our arithmetic on Google’s published rates, not an Ofox charge and not a provider invoice.

Two things in this sample are worth checking on your own traffic before December 1:

  1. On these two easy prompts, Argon high produced about 9× the billable output of medium, and both levels answered correctly. If you run Argon or Carbon at high by default, the thinking level is the first setting to measure. How much it saves on harder prompts, where medium may fail, is something this test cannot tell you.
  2. On these prompts, Hydrogen was not cheaper per request than Argon medium. Its lower per-token rate was outweighed by roughly 550–730 thinking tokens per answer, even at medium, where Argon reported none. That is a reason to measure Hydrogen’s output count before assuming its lower rate means a lower bill, not evidence that Hydrogen costs more in general.

The limits again: two short prompts, two repetitions each, one day, one route (Ofox’s Gemini-native endpoint). Long agentic coding tasks, where Argon also thinks, may behave very differently. Treat this as a reason to measure, not as a result to plan on.

Your options before December 1

Each option below lowers the per-token rate, the token count, or both. Each also has something you must re-test before you switch.

Option 1: Stay on Carbon or Argon and accept the new rate

When it fits: the model passes quality requirements that cheaper options fail, the API spend is small compared with the value of the output, or you have no time to re-validate before December 1.

Trade-off: the token part of the bill doubles. Nothing else changes: same model, same prompts, no migration risk.

What to do anyway: update budgets, quota alerts and any cost dashboards that hard-code $2/$10. If your monthly spend caps were set for November rates, you will hit them twice as fast.

Option 2: Move some traffic to Gemini 4 Neon ($3 / $15)

Neon’s Standard rates are 25% below the new Carbon/Argon rates on input, output and cache. Google positions it between Hydrogen and Carbon: the choosing-a-model page says it “finishes faster with fewer output tokens than Carbon,” and the Neon announcement says that at its default high level Neon has “a 32% lower cost per task than Carbon.”

Google ties that 32% to the Artificial Analysis (AA) Intelligence Index, but it does not state which Carbon thinking level, which Carbon price period or which AA data snapshot it compared against. We could not reproduce it from AA data fetched on November 19, 2026, which prices Carbon at the December $4/$20 rates:

Neon high ($3.74 per AA task, index 59.6) compared withCarbon cost per AA taskDifference
Carbon high, December prices$7.76 (index 60.0, 37.2 min per task)−52%
Carbon medium, December prices$6.59−43%
Carbon low, December prices$7.19−48%
Carbon high, November prices (our halving of AA’s figure)$3.88−3.5%
Carbon medium, November prices (our halving)$3.29+14%

No combination we tried gives 32%. The November rows assume AA’s cost per task is pure token cost, so halving the rates halves it; that is our arithmetic, not an AA number. AA prices Argon at the current $2/$10: its unsuffixed Argon entry costs $4.97 per task (index 50.4), and the same doubling assumption puts it near $9.95 at December rates. AA labels that entry “Gemini 4 Argon (High)”. At Argon’s default medium, AA lists $4.31 per task (index 47.6), about $8.62 at December rates by the same doubling (our arithmetic). These are aggregate benchmark costs, useful for direction only.

Trade-offs and re-tests:

  • Default thinking differs. Neon defaults to high and adds xhigh; Carbon and Argon default to medium and cap at high. If you switch the model ID without setting thinking_level, you also switch to a deeper default, which can raise output tokens and latency. Neon’s migration notes tell you to set the level explicitly.
  • Hop limits differ. The default output limit is 65,536 tokens per response on all four Gemini 4 models. The difference is per non-streaming foreground hop: 65,536 on Neon and Hydrogen, 32,768 on Carbon and Argon, per Google’s long decode guide. Review max_output_tokens and timeouts.
  • Quality is not interchangeable. Re-run your evaluation set and compare cost per accepted result, not per request. Our Neon vs Carbon comparison covers where each is stronger, and the Neon pricing guide breaks down cost by thinking level.
  • Access. Neon is paid-tier only. As of November 19, 2026, Neon is not available on Ofox, so this option means calling Google’s API directly; our Neon API guide shows the native setup.

Option 3: Move some traffic to Gemini 4 Hydrogen ($1.50 / $7.50)

Hydrogen’s rates are 62.5% below the new Carbon/Argon rates. Google’s guidance calls it the default choice for everyday coding, interactive apps and tool-using agents. It is also on a free tier, and through Ofox as google/gemini-4-hydrogen.

Trade-offs and re-tests:

  • Thinking can erase the rate advantage. Hydrogen defaults to high. In our 24-request test it spent roughly 550–730 thinking tokens on easy prompts where Argon medium reported none, which made it more expensive per request on those prompts. Try Hydrogen at low and medium on your own traffic before concluding it saves money.
  • Capability ceiling. Hydrogen is designed for speed and everyday work. Hard multi-file debugging, proofs or security review are where Google points you to Carbon. Expect quality regressions there and measure them.
  • Mixed sessions. Google documents routing planning turns to Carbon and execution turns to Hydrogen in one multi-turn session. If you mix models in a client-managed history, preserve thought signatures exactly; our missing thought_signature fix covers the failure you will see if you don’t.
  • Rate limits. If you move traffic through Ofox, check the default limits in the rate limits guide before shifting a large share at once. The Hydrogen setup guide covers both Google and Ofox access.

Option 4: Lower the thinking level on Carbon or Argon

The per-token rate doubles no matter what, but you control how many output tokens you generate. Google’s own guidance for every Gemini 4 model is to run evaluations at all levels and “pick the lowest level that meets your quality bar.”

Trade-offs and re-tests:

  • Biggest win if you run high today. In our tiny test, high produced about 9× the billable output of medium on two easy prompts. On hard prompts the gap will differ, and medium may fail where high succeeds.
  • Little to gain below medium on simple traffic. Argon low and medium used about the same tokens in our runs because neither reported thinking on those prompts. Google warns that quality on hard tasks “can drop” at low.
  • Watch retries. A cheaper attempt that fails and retries can cost more per accepted result than one high attempt. Keep max_output_tokens at 32,768 or above on Carbon so thinking does not truncate the answer, as the Carbon guide recommends.

Option 5: Get more cache hits

Cached input stays 95% below ordinary input after the change ($0.20 vs $4.00 on Standard). Carbon and Argon support implicit caching only; explicit cache creation is not supported, so you cannot force hits. Google’s guidance: implicit caching applies to Gemini 4 prompts of at least 4,096 tokens; put the shared system instructions, tool definitions and reference documents at the start of the prompt so repeated requests share a prefix.

Trade-off: caching only reduces input cost. In the medium workload above, raising cache hits from 50% to 80% of input saves 90M × ($4.00 − $0.20) = $342 a month at December rates. That is meaningful, but output was still $600 of the bill. Restructuring prompts also changes what the model sees first, so re-run quality checks. Verify hits from reported cached-token counts, not from the fact that your prompts look repetitive. Agent loops have an extra rule (a turn must add at least 2,050 tokens for the full previous request to be cached), explained in the Carbon task-cost guide.

Option 6: Move non-urgent work to Batch or Flex

This is the only lever that keeps the per-token rate where it is today. December Batch and Flex rates ($2 / $10, cached $0.10) are exactly the current Standard rates. In the table above, every workload’s December Batch bill equals its November Standard bill.

Trade-offs and re-tests:

  • Batch is asynchronous, with a target turnaround of up to 24 hours. It suits evaluations, dataset processing, backfills and nightly reports. It does not suit anything a user is waiting on.
  • Flex is synchronous, but best-effort with a 1–15 minute target, and requests can be preempted during traffic spikes. Your client needs timeouts and retry handling built for that. Our Argon API error guide and the general 429 retry guide cover backoff patterns.
  • Grounding prices are listed on the Standard and Priority tabs but not on Batch or Flex for these models, so check that your grounded requests are supported before moving them.
  • Priority is the opposite move. It doubles to $7.20 / $36. If you use Priority for latency, confirm which requests actually need it.

Combining options: a blended bill

Most teams will mix these. To estimate a mix, give each route a traffic share s and an output multiplier k measured from a pilot (output tokens on the new route divided by output tokens on Argon or Carbon for the same requests):

blended = Σ over routes:  s × [ (I − C) × Pi + C × Pc + k × O × Po ]

Example, medium workload at December rates, with an assumed k = 1.5 for Hydrogen (an illustration, not a measurement): keep 60% on Argon Standard (0.6 × $1,230 = $738) and move 40% to Hydrogen (0.4 × ($225 + $11.25 + 1.5 × 30 × $7.50) = $229.50). The total is $967.50, against $1,230 if nothing moves. With k = 6, close to what our easy-prompt test showed against Argon medium, the Hydrogen share alone would cost $634.50 and the total would rise to $1,372.50. Measure k on your own traffic before you commit.

Decision checklist to finish before December 1

  1. Pull November usage by model, tier and thinking level, including reported cached tokens and thinking tokens. Without these four numbers you cannot recompute anything.
  2. Recompute December with the formula above. Doubling the token part works only if tier and mix stay fixed; add grounding separately.
  3. Find your thinking_level settings. Anything on high is the first place to test medium.
  4. Tag traffic by urgency. Everything that can wait minutes goes to the Flex shortlist; everything that can wait hours goes to Batch.
  5. Build a small evaluation set (20–50 real requests with pass/fail criteria) and run it on your current model, on a lower thinking level, and on Neon and Hydrogen with thinking_level set explicitly.
  6. Compare cost per accepted result, including retries, not cost per request. Our cost-per-correct-result guide shows how to set that up.
  7. Check prompt structure for caching. Is the stable part at the start and above 4,096 tokens?
  8. Update budgets and alerts to December rates before the first week of December spend comes in.
  9. Avoid the boundary. Do not schedule large jobs on November 30 expecting the old rate; the time zone is not documented.
  10. Re-check the pricing page on December 1. Rates and qualifiers can change; this article reflects the November 19 page.

For model fit beyond cost, use our Gemini 4 model selection guide.

If you call Argon through Ofox

Gemini 4 Argon is listed on Ofox as google/gemini-4-argon (checked against the authenticated model list on November 19, 2026). Its model page lists the supported protocols (OpenAI-compatible and Gemini native) and the price that Ofox charges. We do not quote or predict Ofox prices here. Check the Argon model page and the Ofox pricing and billing docs directly before and after December 1.

Gemini 4 Carbon and Gemini 4 Neon are not available on Ofox as of November 19, 2026; use Google’s API for those. Hydrogen is available as google/gemini-4-hydrogen. The thinking-level and caching options above work the same way on any route that passes thinkingConfig through; on Ofox, the native endpoint is described in the Gemini API reference, and per-request token and cost breakdowns appear in usage tracking.