Gemini 4 Carbon vs Argon: which Gemini 4 model should you use?

Use Gemini 4 Argon for interactive coding and fast loops, and move a task to Gemini 4 Carbon when Argon's answer misses something you care about. That is also Google's own advice: the two models take the same request, have the same 1,048,576-token context and launched at the same introductory price. What differs is how much Carbon thinks before it answers.

Most comparisons you will find were written for the October 28 announcement, when only a small group of cybersecurity defenders could use Carbon. That changed on November 10, 2026: Google opened Carbon to every paying Gemini API developer. So we ran it, and then ran the same prompts on Argon.

What is the same

Google's migration guide is blunt about it: if you already run gemini-4-argon, "upgrading to gemini-4-carbon requires no request schema changes". You change the model string and nothing else.

Both ignore temperature, topP and topK, and neither lists a minimal thinking level (Carbon returned HTTP 400 when we sent one). If your code works with one, it sends a valid request to the other.

What Google says is different

Google describes Carbon as the model "built for tasks that push the limits of reasoning depth". In the same guide it says Carbon "spends more thinking tokens per turn to explore thoroughly and verify its work, which increases per-turn token use and latency compared to Argon". Google points Carbon at root-cause debugging, architecture work, dense multimodal synthesis, proofs and security audits. Argon gets "interactive coding, day-to-day feature work, and fast production loops".

The rest of the gap comes from Google's own announcement. It lists stronger cybersecurity results, fewer unsupported claims and better agentic honesty for Carbon. We have not reproduced those benchmarks, so we will not repeat the numbers as if we had. Read them in Google's Carbon announcement with that in mind.

Same prompt, both models: what we measured

On November 11, 2026 we sent three prompts with known answers to both models through SeedRouter's native Gemini endpoint. Every request used thinkingLevel: "medium" and maxOutputTokens: 8192, with one attempt each and no retries. Output counts the visible answer; thinking is what thoughtsTokenCount reported.

Three things stood out.

Carbon did not always think more. On the factorial question Argon spent 249 thinking tokens and Carbon 71. On the other two, Argon answered with no reported thinking at all, and Carbon thought for 224 and 392 tokens. Across the three prompts Carbon billed 1,010 output tokens (answer plus thinking) against Argon's 701. That is about 44% more on this tiny sample, which matches Google's direction but not its every case.

Thinking did not change the answers here. Both got all three right. Both rewrote the function the same way: sort descending, then slice [:k]. Neither pointed out that the original ranked[-k:] also misbehaves when k is 0, because -0 is just 0 and the slice returns the whole list. The corrected code happens to handle it. Easy prompts like these cannot show where Carbon's extra checking pays off; they only show what it costs.

Latency followed thinking, not model name. Carbon's slowest answer was the code fix (18.4 seconds), the one where it thought longest. On the factorial prompt it was faster than Argon.

Three prompts prove nothing about quality. They do tell you how to run your own test: take five or ten tasks where Argon currently fails or needs a second pass, send them to both models with the same settings, and compare correct answers per dollar, not tokens per answer.

Cost: same price per token, different tokens per answer

Both models launched at the same introductory Gemini API price, so the difference you pay comes from token counts, mostly thinking. Google lists these introductory rates per 1M tokens (checked on November 11, 2026), valid through November 30, 2026:

These are Google's list prices from its pricing page. Current per-token rates on our platform are below, and they apply equally to both models in the same group:

Because thinking is billed as output, a task where Carbon thinks 400 tokens and Argon thinks none costs more on Carbon even at identical prices. If you move a whole workload to Carbon, budget for higher output usage and set maxOutputTokens high enough that thinking does not cut the answer short. Google suggests at least 32,768.

How to switch a request from Argon to Carbon

Change the model in the URL:

curl "https://api.seedrouter.ai/v1beta/models/gemini-4-carbon:streamGenerateContent?alt=sse" \
  -H "Authorization: Bearer $SEEDROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"contents":[{"role":"user","parts":[{"text":"Review this migration plan for ordering mistakes."}]}],"generationConfig":{"thinkingConfig":{"thinkingLevel":"medium"},"maxOutputTokens":32768}}'

Two practical notes from our production tests. First, use streaming for anything that might run long. A non-streaming audio request of ours ran for about two minutes and ended with HTTP 524 at the edge before the answer arrived; the same request streamed fine in 84 seconds. Second, keep thinkingLevel at medium unless a task needs high. In our checks, high thinking on a small factoring question used 493 thinking tokens against 193 at medium, with the same answer.

The full request reference is in the Gemini 4 Carbon API docs, and you can try both models with your own prompts on the Gemini 4 Carbon page and the Gemini 4 Argon page.

FAQ

Is Gemini 4 Carbon better than Argon?

On Google's benchmarks, yes, especially for long agentic work and security tasks. On our three simple prompts, both answered correctly, and Carbon used more output tokens on two of them. Better shows up on hard tasks; test the ones Argon gets wrong.

What is the difference between Gemini 4 Carbon and Argon?

Thinking effort. The request format, context window, tools, modalities and introductory price are the same. Carbon thinks more per turn to verify its work, so it is slower and uses more output tokens on hard prompts.

Is Gemini 4 Carbon more expensive than Argon?

Not per token: both launched at $2 input and $10 output per 1M tokens on Google's list. Per answer it usually costs more, because Carbon spends more thinking tokens, which are billed as output.

Can I switch from Gemini 4 Argon to Carbon without code changes?

Yes. Google's guide says no request schema changes are needed. Change gemini-4-argon to gemini-4-carbon and review your output limit and timeouts.

Does Gemini 4 Carbon replace Argon?

No. Google keeps Argon as the faster option for most software and knowledge work and positions Carbon for the hardest problems. Both stay available.