Gemini 4 Carbon vs Gemini 4 Argon
Model pages: Gemini 4 Carbon API · Gemini 4 Argon API — pricing, specs and the playground
The choice between Carbon and Argon is not about price, because there is no price difference per token: Google lists both at $2 / $10 per 1M tokens through 30 November 2026 and $4 / $20 after. It is about how much thinking you want to pay for and wait for on each request.
Google’s documentation puts it plainly: Carbon spends more thinking tokens per turn, which raises token use and latency, and suits complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs and security auditing. Argon stays the recommended default for interactive coding, day-to-day feature work and fast production loops.
Every number below comes from public sources read on 12 November 2026: Google’s own announcements and documentation, and independent evaluations from Artificial Analysis, Vals, Cursor, Zapier, Arena and one developer’s hands-on test. We have not run our own benchmark.
Gemini 4 Carbon
More thinking, higher scores, slower start
- ·Artificial Analysis Intelligence Index 60 (Argon 50)
- ·CursorBench 4.0 60.4% vs 52.9%; Terminal-Bench 4.0 65.2% vs 57.1%
- ·About 68 s to the first answer token (Argon about 38 s)
- ·Google: for root-cause debugging, architecture, proofs and security audits
Gemini 4 Argon
Google’s default for most work
- ·Google: the ideal choice for most software engineering and knowledge work
- ·Wins Vals’s Vibe Code Bench (91.91% vs 89.06%) and HLAB
- ·Fewer thinking tokens per turn, so faster replies
- ·Same price per token, same 1M context and inputs as Carbon
Default to Gemini 4 Argon for interactive coding and everyday work. Send a task to Gemini 4 Carbon when Argon gets stuck, when the answer has to be right the first time, or when the job is root-cause debugging, architecture, very long context or a security review.
What is the same
Carbon is not a different family. Google’s model card says Gemini 4 Carbon is based on Gemini 4 Argon, and the API documentation lists the same context window of 1,048,576 tokens shared by input and output, the same 65,536-token default output limit (up to 1,048,576 with long decode), the same inputs (text, image, video, audio and PDF, text out) and the same thinking levels, low, medium and high, with medium as the default.
Moving from Argon to Carbon needs no schema change: Google’s migration notes say upgrading to gemini-4-carbon only changes the model string. Two Carbon-specific details: temperature, top_p and top_k are ignored, and only implicit caching is available (prompts of 4,096 tokens or more), with no explicit cachedContents.
Google’s own numbers
In Google’s launch table, run at the highest thinking setting, Carbon beats Argon on every row. The gaps are widest on long-context reasoning (GraphWalks 96.2% against 84.2%), science (Terminal-Bench-Science 64.8% against 52.9%) and the newest coding tests. Google also reports that Carbon passed 92% of an agentic honesty test against Argon’s 70%, and cut unsupported claims by about half.
| Benchmark (Google, highest thinking) | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| SWE-sweep | 9.6% | 4.3% |
| FrontierSWE 2 | 61.2% | 55.0% |
| Terminal-Bench 4.0 | 65.2% | 57.1% |
| OSWorld-2.1 (offline subset) | 81.5% | 74.7% |
| AutomationBench (Zapier) | 53.9% | 51.3% |
| GraphWalks, 256K to 1M | 96.2% | 84.2% |
| RiemannBench | 80.0% | 76.0% |
| Terminal-Bench-Science 0.1 | 64.8% | 52.9% |
| Chartography | 72.5% | 71.6% |
| Agentic honesty (375 tasks) | 92% | 70% |
Independent benchmarks
Third-party tests mostly agree with Google, but the margin is smaller and Argon wins a few. Artificial Analysis scores Carbon 10 points higher; Cursor’s and Zapier’s own benchmarks rank Carbon first; on Vals the two are about a point apart and Argon wins five of the sub-benchmarks, including Vibe Code Bench. On Arena’s text leaderboard they are statistically tied, while on WebDev Arena Carbon ranks 4th and Argon 12th.
| Test | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Artificial Analysis Intelligence Index (High) | 60 | 50 |
| CursorBench 4.0 (High) | 60.4% | 52.9% |
| Zapier AutomationBench (High) | 53.88% | 51.29% |
| Vals Index v2.2 | 70.45% (3rd of 36) | 69.39% (5th) |
| Vals Vibe Code Bench v1.1 | 89.06% | 91.91% |
| Vals HLAB-Vals-1.1 | 17.50% | 27.08% |
| Arena text leaderboard | 1526 (preliminary) | 1522 |
| WebDev Arena | 1768, 4th (preliminary) | 1682, 12th |
| paddo.dev: runs passing every test (6 changes × 20) | 17 | 16 |
A hands-on test on real code
The most useful public test so far comes from paddo.dev, published on 11 November 2026. The author took six changes already merged in a production monorepo, ran each 20 times per model at medium effort, and graded the runs with the original fixes’ tests. Carbon passed every test in 17 runs, Argon in 16, and Claude Fable 5.2 and Opus 5.5 in 13 each.
The cost of that edge was time and output: Carbon wrote 1.82M output tokens to Argon’s 1.37M and took 30 minutes per fix on average, the longest 46. The author calls it a tie with Argon, and notes that the two did not run in the same harness (Argon in Claude Code, Carbon in Antigravity CLI).
Speed and cost per task
Carbon starts slower but does not always finish later. Artificial Analysis measured about 68 seconds to Carbon’s first answer token against 38 for Argon, and a longer average single request (159 seconds against 111). On its multi-step agent tasks, though, Carbon averaged 25.5 minutes against Argon’s 30, at 80 output tokens per second against 65.
Cost depends on which test you look at. Per Artificial Analysis task, Carbon averaged $7.76 at list price against Argon’s $9.95, about $2.13 against $2.74 at apimodels.app prices through 30 November and about $4.27 against $5.47 from 1 December. Running Artificial Analysis’s whole index cost more on Carbon, $8,240 against $7,620, with 205M output tokens against 113M. On CursorBench, Carbon’s cost per task was about twice Argon’s ($9.22 against $4.57). Budget for Carbon to cost the same or more per job, never less per token.
| Artificial Analysis (both at High) | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Time to first answer token | 67.83 s | 37.64 s |
| Output speed | 80 tok/s | 65 tok/s |
| Single request, end to end (mean) | 159.39 s | 110.88 s |
| Time per index task | 25.5 min | 30.0 min |
| Cost per index task (list price) | $7.76 | $9.95 |
| Cost to run the whole index (list price) | $8,240 | $7,620 |
How to switch between them
On apimodels.app both models sit behind the same key and the same endpoints, so switching is the model field: gemini-4-argon or gemini-4-carbon. On the OpenAI-compatible /v1/chat/completions endpoint, reasoning_effort low, medium or high maps to Gemini’s thinking level; on the native /v1beta generateContent endpoint, set generationConfig.thinkingConfig.thinkingLevel directly.
Give Carbon room to think. Google recommends a max output of at least 32,768 tokens, because the limit covers thinking plus the answer, and HTTP timeouts of at least 15 minutes. For long Carbon runs, stream the response instead of waiting for one large reply. Lowering Carbon to the low thinking level is a reasonable middle ground when Argon’s answers are close but not quite right.
A common setup is to route by task: Argon for chat, quick edits and high-volume jobs, Carbon for the tickets Argon fails and for reviews where a mistake is expensive. For short, high-volume requests neither is the right pick; Gemini 3.8 Flash costs $0.45 / $2.25 per 1M tokens on apimodels.app.
Caveats
Google’s table compares Carbon with Claude Fable 5.1, not 5.2, and was run at the highest thinking setting; third-party tests mostly ran both models at High. Carbon’s Arena scores are marked preliminary. Artificial Analysis prices both models at $4 / $20, Google’s list price from 1 December; our per-task figures scale those by our own prices. Carbon’s model card also warns of occasional slowness or timeouts at higher thinking levels.
Sources
Read on 12 November 2026. Google: the Gemini 4 Carbon announcements of 28 October and 10 November, the Gemini API docs and the Carbon model card. Third party: Artificial Analysis, Vals AI, Cursor, Zapier, Arena and paddo.dev. apimodels.app prices are our own.
Quick start
cURL
curl https://api.apimodels.app/v1/chat/completions \
-H "Authorization: Bearer $APIMODELS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-4-carbon",
"reasoning_effort": "high",
"stream": true,
"max_tokens": 32768,
"messages": [{"role": "user", "content": "This test fails one run in fifty. Find the root cause: ..."}]
}'
# same key, same endpoint: change "model" to "gemini-4-argon" for everyday workFAQ
What is the difference between Gemini 4 Carbon and Gemini 4 Argon?
Gemini 4 Carbon spends more thinking tokens per turn than Gemini 4 Argon; that is the main difference, according to Google. Carbon is built on Argon and shares its 1M-token context, its inputs (text, image, video, audio, PDF), its thinking levels and its per-token price. The extra thinking gives Carbon higher scores on most tests, 60 against 50 on the Artificial Analysis index, and makes it slower to start, about 68 seconds to the first answer token against 38.
Is Gemini 4 Carbon better than Argon?
Gemini 4 Carbon scores higher on most benchmarks, but not on all of them, and Google still recommends Argon for most work. Carbon leads on Artificial Analysis (60 vs 50), CursorBench 4.0 (60.4% vs 52.9%) and every row of Google’s own table. Argon wins Vals’s Vibe Code Bench and HLAB, ties Carbon on Arena’s text leaderboard, and in one hands-on test on real code passed 16 of 20 runs against Carbon’s 17. Use Carbon where a better answer is worth the wait.
Which costs less, Gemini 4 Carbon or Argon?
Neither is cheaper per token: both cost $1.10 per 1M input and $5.50 per 1M output on apimodels.app through 30 November 2026, and $2.20 / $11.00 from 1 December, 45% below Google’s list. Per job, Carbon usually costs the same or more because it writes more thinking tokens. Artificial Analysis found Carbon slightly lower per agent task ($7.76 vs $9.95 at list price) but higher for its whole index ($8,240 vs $7,620), and Cursor measured about twice Argon’s cost per task.
Which is faster, Gemini 4 Carbon or Argon?
Gemini 4 Argon replies sooner: Artificial Analysis measured about 38 seconds to Argon’s first answer token against 68 for Carbon, and 111 seconds against 159 for an average single request. Carbon generates faster once it starts (80 against 65 tokens per second) and finished Artificial Analysis’s multi-step agent tasks in 25.5 minutes against 30. For chat and interactive coding, Argon feels faster.
How do I switch from Gemini 4 Argon to Carbon?
Change the model field from gemini-4-argon to gemini-4-carbon; Google’s migration notes say no other request change is needed. Then give Carbon room: raise the output limit to at least 32,768 tokens, set HTTP timeouts to 15 minutes or more, stream long responses, and drop temperature, top_p and top_k, which Carbon ignores. On apimodels.app the key and endpoints stay the same.
Can I use Gemini 4 Carbon and Argon from mainland China?
Not through Google directly: Google’s list of regions for the Gemini API and AI Studio does not include mainland China, Hong Kong or Macao. Through apimodels.app both models work from mainland China without a proxy: create a key, top up with Alipay, PayPal, a card or USDT, and call gemini-4-carbon or gemini-4-argon on api.apimodels.app.
Related
- Gemini 4 Carbon model page
- Gemini 4 Argon model page
- Claude Fable 5.2 vs Gemini 4 Carbon
- How to use Gemini 4 Carbon
- Gemini 4 Carbon case library: 223 credited cases
- Gemini 3.8 Flash model page
- LLM API docs
