Gemini 4 Carbon vs Argon: same-length answers, three times the wait
The short version
- On 43 shared prompts, Gemini 4 Carbon and Gemini 4 Argon wrote answers of the same length: 385 and 386 words on average.
- Carbon was slower on all 43. Median time to a full answer was about 35 seconds against about 11.
- Carbon used 2,006 output tokens per answer on average, reasoning included, against 1,099 for Argon. The same question costs more on Carbon.
- On easy prompts the two answers were nearly identical. The gap opened on a hard SQL optimization problem, where Carbon found a bug Argon missed.
- Google's own advice (Argon for everyday work, Carbon for the problems Argon can't crack) matches what we measured.
Google opened Gemini 4 Carbon to all developers on November 10, 2026, and calls it its strongest model. In the same post it says Argon is still the better fit for most software engineering and knowledge work. We ran both through FastMetal's shared prompt set (21 in English, 22 in Japanese) to see where the difference actually lands.
How we measured
FastMetal keeps 43 shared prompts (summarization, bug fixes, translation, word problems and more) and publishes every model's answers on its model page. Each model got each prompt once, and we recorded the answer, the time to a complete answer, and the output token count.
Argon's answers were recorded on October 13, 2026, and Carbon's on November 10. Times include network time and load on different days, so treat the timing gap as indicative.
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Average answer length | 385 words | 386 words |
| Median time to answer | 34.9 s | 10.5 s |
| Average output tokens (incl. reasoning) | 2,006 | 1,099 |
| Prompts where Carbon was slower | 43 of 43 |
The visible answer is the same size. What changes is how much the model thinks first: Carbon used more output tokens on 38 of the 43 prompts, which fits Google's description of a model that explores, reasons and verifies more than Argon. Reasoning tokens bill as output, so per-request cost is higher on Carbon. Rates are on the Gemini 4 Carbon and Gemini 4 Argon model pages.
Easy prompts: the same answer
Both counted the letters of "strawberry" one by one and found three r's.
On a Japanese word problem about two machines and their defect rates, all three answers matched (1,610 good units, about 13 h 10 min, about 11 h 58 min). The one difference: Argon's Japanese answer slipped a Korean word into a sentence, "방程式" for the Japanese 方程式 (equation). Carbon's answer was clean, and that was the only answer from either model with Korean in it.
On a Python TypeError fix, both answers followed the same reasoning, down to recommending Decimal for money.
A hard prompt: SQL optimization
The prompt: a five-table query over 5 million rows takes more than 30 seconds; optimize it. Carbon spent 104 seconds and 6,615 output tokens on it. Argon took 12 seconds and 1,248 tokens.
Both caught the main bug: joining orders and reviews side by side multiplies rows and inflates the sales totals. Both moved the aggregation into separate CTEs. Carbon then went further:
- It picked the top 50 products by sales first, and only then joined categories and averaged reviews for those 50 rows instead of every product.
- It pointed out that
o.amountis the whole order's amount, so an order with several line items adds the full order total to each product. For per-product sales it suggests summing the line-item amounts. Argon's answer doesn't mention this. - Its indexes include the aggregated columns, so the totals can come from the index alone.
On this prompt the extra thinking turned into concrete findings.
Which to use
- Everyday work: Argon. For writing, summarizing, translation and ordinary code fixes, the answers barely differed, and Argon is faster and cheaper per request.
- Hard problems: Carbon. When the job is to question assumptions, redesign something or find a hidden bug, Carbon's extra thinking showed up in the answer. Escalating to Carbon when Argon's answer falls short wastes the least.
- Building an agent: both are on the FastMetal API and MCP server, as
gemini-4-argonandgemini-4-carbon.
FAQ
Is Gemini 4 Carbon smarter than Argon?
Google calls Carbon its strongest model. On the LMArena dataset published November 9, 2026, Carbon was first on the overall text board and Argon second, with the scores inside the margin of error. On our 43 prompts the easy answers were nearly identical, and the difference showed on the hard SQL problem.
Why is Carbon slower?
It thinks more before answering. Its answers are the same length as Argon's, but it used about 1.8 times the output tokens on average. Google says the same: Carbon spends more tokens to explore and verify.
Does it cost more?
Both bill per million input and output tokens, at different rates, and Carbon spends more output tokens on the same prompt, so each request costs more. Since November 11, 2026, Carbon's FastMetal rate follows the price Google applies after its introductory period. Current rates are on each model page.
Can I call Carbon from the API?
Yes. Since November 11, 2026, Gemini 4 Carbon is on the API and MCP server, like Gemini 4 Argon, as gemini-4-carbon.
Bottom line
Carbon writes the same amount as Argon after thinking a lot longer. On easy prompts that extra thinking changes nothing; on the hard SQL problem it found a real bug. Use Argon by default and hand Carbon the problems Argon can't finish.
All 43 answer pairs are on the Gemini 4 Carbon and Gemini 4 Argon model pages. To try both, switch models in chat.