Note the date before you read the price. Gemini 4 Carbon is on introductory pricing of $2 per million input tokens and $10 per million output through 30 November 2026. From 1 December the rate doubles to $4 and $20. Any cost model built this month needs the second number.
Carbon is the top of Google's Gemini 4 line: the model for long, multi-step agentic work, complex coding, open-ended scientific research and cyber defence. It sits above Gemini 4 Argon, which launched two weeks before Carbon's first release.
A restricted start, then general release
Two dates circulate, and both are right. Carbon was announced on 28 October 2026 and first released to a subset of cyber defenders through Google's Fairwind Program. General availability on the paid tier of the Gemini API followed on 10 November. Use 10 November as the date you could build on it; 28 October as the date it existed.
| Field |
Value |
| API ID |
gemini-4-carbon |
| Status |
Generally available (paid tier only) |
| Context window |
1,048,576 tokens |
| Input |
Text, image, audio, video, PDF |
| Output |
Text |
| Max output |
Stated as "64K / 1M" across Gemini 4; catalogues list 65,536 |
| Reasoning |
Yes |
| Free tier |
No |
The output limit is unclear. Google gives "64K / 1M" for the whole Gemini 4 family without splitting it by model, and third-party listings record 65,536. Set max_output_tokens explicitly and test a long generation before you rely on anything past 64K.
The token bill is the real price
Google is open about this: Carbon spends more tokens than Argon on the same task, because it explores more, reasons further and runs longer verification loops. The per-token rate is the same as Argon during the introductory window. The per-task cost is not.
| Tier |
Until 30 Nov 2026 |
From 1 Dec 2026 |
| Standard (in / out) |
$2 / $10 |
$4 / $20 |
| Batch or Flex |
$1 / $5 |
$2 / $10 |
| Priority |
$3.60 / $18 |
$7.20 / $36 |
| Implicit caching |
$0.10 |
$0.20 |
Search grounding and Maps grounding each include 5,000 free requests a month, then $14 per thousand. All indicative; check your billing console, especially after the December change.
What Google reports
Publisher-reported, not independently checked: first place on the Artificial Analysis Intelligence Index v5.0 with a score of 59, SEC-bench Pro at 88.2% for finding vulnerabilities, ExploitGym at 54.8% and CWE-bench at 71.0% for patching. Google also says internal teams have seen it run for days on root-cause-and-patch work in large codebases.
Carbon or Argon?
Google's own recommendation is blunt. Argon is the default for most software engineering and knowledge work. Carbon is for the harder slice: tangled multi-step debugging, and research questions without a known answer. A team moving up from Gemini 3.8 Flash should usually land on Argon first.
Test it the way you'll pay for it
Pick ten tasks where Argon gives up or gets it wrong. Run them on Carbon and log tokens per task, not just pass rate. Then reprice the results at the December rate. If Carbon solves four of the ten and the per-task cost stays under what an engineer's hour costs you, route that category up. Otherwise keep it as a manual escalation.
Beyond the API, Carbon is rolling out to Google AI Ultra subscribers in Antigravity, including the /swarm preview that coordinates teams of agents, as well as in the Gemini app and AI Mode in Search. It is also in Android Studio and Gemini for business. More Google models are on the provider page.