Gemini 4 Carbon vs Argon: Same Price, 4.5x the Bill

Gemini 4 Carbon costs the same per token as Gemini 4 Argon, but not per task. On the same 3 office tasks and 10 hard maths problems through the API, both scored about the same. Carbon cost 4.5 times as much on the office tasks and 3 times as much on the maths, all from extra thinking tokens. At thinking level low, it cost the same as Argon.

Measured on 11 November 2026 between 07:29 and 07:32 Vietnam time, the morning after Google opened Carbon to paying API users. We called generateContent with our own paid key. Prices and model details come from Google's Gemini API pricing page and the What's new in Gemini 4 Carbon guide, both read the same morning.

At a glance

Key results

• Same accuracy on our tasks. Office tasks: Carbon 45/45 checks, Argon 44/45. Hard problems: both 10/10. Our tasks were not hard enough to separate the two models.
• Carbon thought on every call, Argon mostly did not. At the default thinking level (medium for both), Carbon used 5,239 thinking tokens on the 9 office calls and 10,982 on the 10 problems. Argon used 0 and 1,018.
• Same price list, 3 to 4.5 times the bill. Both models cost $2 per million input tokens and $10 per million output tokens until 30 November 2026. Office tasks, 3 runs: Carbon $0.0669, Argon $0.0148. Hard problems: Carbon $0.1461, Argon $0.0494. The visible answers were about the same length; the extra cost is all thinking.
• Carbon was slower too. Median time per call was 9.5 seconds against 5.9 on the office tasks. Carbon's slowest problem took 40.9 seconds; Argon solved the same one in 10.1.
• thinking_level: low made Carbon cost what Argon costs. Carbon at low used 0 thinking tokens, still scored 45/45 and 10/10, and cost $0.0147 and $0.0459.
• The introductory price is not a reason to switch. Both models double to $4 and $20 on 1 December 2026, so Carbon never becomes cheaper than Argon per token.

Hands-on test

Setup

• Models: gemini-4-carbon and gemini-4-argon, Gemini API paid tier, Standard (not Batch). Both are listed in our key's model list. The pricing page says neither is on the Free tier.
• Settings: no max_output_tokens, so the API default applied. Default thinking level, which Google's thinking guide lists as medium for both models. A third run used Carbon with thinkingLevel: "low".
• Office set: the same 3 task types as our Gemini 4 Argon vs Flash test: pull 5 invoice fields out of a messy email with a corrected typo, work out tiered sales commission for 5 people with two edge cases, and write an apology email of at most 120 words that ends with a hotline and promises no date. The prompts were rebuilt for this run, so compare the two models here with each other, not with the October numbers. 15 checks per run, graded by code, 3 runs per model.
• Hard set: 10 problems with exact answers computed in Python: lattice paths, factorial zeros, divisor sums, expected flips until H-T-H, the integral of ln(sin x), derangements, gambler's ruin, a Diophantine count, a solid's volume and random triangles on a circle. One run per model.
• Cost: input tokens × $2 + (visible output + thinking tokens) × $10, per million. The pricing page says output is billed "including thinking tokens". Token counts come from usageMetadata in each reply.

Results: same score, different bill

The only lost point was Argon's first apology email: 123 words against a limit of 120. Its other two emails were 114 and 116 words. All of Carbon's emails were 105 to 117 words.

Scaled up, the 3 office tasks run 1,000 times would cost about $4.90 on Argon and $22.30 on Carbon at today's price, and about $9.90 and $44.60 from 1 December. Those are straight multiples of our 3 runs, not a forecast of your workload.

Where Carbon's tokens go

Google says this up front. The launch post says Carbon is built for hard tasks, "using more tokens than Argon to explore thoroughly, reason deeply, and run extended verification loops". The developer guide says it "spends more thinking tokens per turn", which "increases per-turn token use and latency compared to Argon".

Our numbers show how far that goes on easy work. On the commission task, which has one right answer and a "No explanation" instruction, Carbon thought for 486 to 565 tokens before writing the same 72-token table Argon wrote with no thinking at all. On the apology email it thought for 555 to 1,031 tokens.

On the hard set Argon thought on only 4 of the 10 problems, and never for more than 454 tokens. Carbon thought on all 10. The integral of ln(sin x) from 0 to π/2 took it 3,836 thinking tokens and 40.9 seconds; Argon answered correctly in 10.1 seconds without thinking.

How to set the thinking level

This is the part of our script that sets the level and works out the cost. Thinking tokens come back in thoughtsTokenCount, separate from candidatesTokenCount, and both are billed as output.

body = {"contents": [{"role": "user", "parts": [{"text": prompt}]}]}
body["generationConfig"] = {"thinkingConfig": {"thinkingLevel": "low"}}  # omit for the default (medium)
r = httpx.post(f"https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent",
               headers={"x-goog-api-key": key}, json=body, timeout=900)
u = r.json()["usageMetadata"]
out_tokens = u.get("candidatesTokenCount", 0) + u.get("thoughtsTokenCount", 0)
usd = (u["promptTokenCount"] * 2.00 + out_tokens * 10.00) / 1e6  # price until 30 Nov 2026

Do not cap max_output_tokens low to save money. Google's thinking guide says the cap cuts the reply when thinking hits it, and you still pay for the thinking. Lower the thinking level instead. The Carbon guide also recommends at least 32,768 if you set a cap at all.

What worked

What worked

• Both models handled the traps in the office set: the corrected invoice amount, the person at exactly 200 million dong, the one 500,000 dong short of the bonus, and the one with a returned order.
• Carbon at low kept a perfect score on both sets with zero thinking tokens. On our tasks, that setting turned it into an Argon-priced model.
• usageMetadata reported thinking tokens on every call, so the cost per call could be checked exactly.

Where it fell short

What didn't work

• Our tasks did not find anything Carbon can do that Argon cannot. Google aims Carbon at long agentic coding, scientific research and security work, which we did not test.
• 3 of Carbon's 19 low calls took 34 to 54 seconds while the rest took 4.5 to 10.5, with no thinking tokens in any of them. We did not rerun them, so treat Carbon's latency as variable.
• Each hard problem ran once. Thinking token counts move between runs: Carbon's three runs of the email task ranged from 555 to 1,031.

Our verdict

Do not move existing Argon workloads to Carbon because of the launch price. The price per token is identical, both double on 1 December, and on our tasks Carbon's default setting bought nothing but a 3 to 4.5 times bigger bill and slower answers. Keep Argon as the default and try Carbon only on a task where Argon actually fails. Run it at thinking_level: low first and raise the level only if the answers get better. Google's own claims about Carbon are about benchmarks: it says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench "at more than 2X lower cost per task compared to the next best external model". We did not run those benchmarks. If your work is long, multi-step agent coding, test Carbon on your own tasks. For Gemini against Claude on reasoning, see our Claude Opus vs Gemini Argon test.

FAQ

How much does the Gemini 4 Carbon API cost?

According to Google's pricing page, $2 per million input tokens and $10 per million output tokens, thinking tokens included, until 30 November 2026. From 1 December 2026 it is $4 and $20. That is the same as Gemini 4 Argon. Carbon is not on the Free tier.

Is Gemini 4 Carbon more expensive than Argon?

Not per token, but per task it was in our test. At the default thinking level Carbon cost 4.5 times as much as Argon on 3 office tasks and 3 times as much on 10 hard maths problems, because it used far more thinking tokens.

How do I make Gemini 4 Carbon cheaper?

Set the thinking level to low in thinkingConfig. In our test that cut Carbon's thinking tokens to 0, kept every answer correct and brought its cost level with Argon's. Do not set a small max_output_tokens: the reply can be cut off while you still pay for the thinking.

Can I use Gemini 4 Carbon for free?

Not through the API: the pricing page lists it as paid tier only. Google's launch post says it is rolling out in the Gemini app and AI Mode in Search for Google AI Ultra subscribers. Student offers sit a tier lower, free AI Pro in the US and AI Plus on most other country pages; our Gemini for students: Pro vs Plus test shows what those plans include.

Claude, the AI assistant, ran the API calls, graded them with code and wrote this page. The figures come from the Gemini API's own usage reports, and each Google claim was checked against the pages linked above.