Google's Gemini 4 Carbon is out, but only for $100-a-month subscribers and developers
Google's strongest AI model opened up on November 10, but regular Gemini users won't see it unless they pay for the top Google AI Ultra plan.
- Google made Gemini 4 Carbon, its most powerful AI model, generally available on November 10, 20261,6.
- In the Gemini app and Google Search it's only for Google AI Ultra subscribers, a plan that starts at $100 a month in the US1,5.
- Independent tester Artificial Analysis ranks it near the top, now just behind Anthropic's Claude Fable 5.28,10.
- Developers pay $2/$10 per million tokens until November 30, then $4/$201.
Should you care?
2/5 Matters if you pay for Ultra or code
Hype check
Benchmark charts are loud, but most Gemini users stay on Gemini 4 Argon or lower, and Google still points most coding and knowledge work to Argon7.
We'll check back on Feb 14, 2027: did it live up to the hype?
What it costs where you live
Who is this for?
- ✕Free Gemini app usersSkip itCarbon is not offered on the free tier; in the app it is limited to AI Ultra subscribers1.
- ✕Google AI Pro subscribersSkip itPro keeps Gemini 4 Argon, and Google has given no timeline for Carbon on cheaper plans7.
- ✓Google AI Ultra subscribersFor youCarbon is rolling out to you in the Gemini app, AI Mode in Search and Antigravity1.
- ✓Developers building with AI APIsFor youIt is in the paid Gemini API at an introductory $2/$10 per million tokens through November 301.
- ~Security teamsMaybeA cyber version with fewer guardrails is available, but only to approved members of Google's Fairwind Program1,5.
Google's smartest AI is here, but it's behind the most expensive door
Google released Gemini 4 Carbon, its most powerful AI so far. Unless you pay $100 a month or write software, your Gemini app won't change.
Google makes an AI assistant called Gemini. It lives in the Gemini app on Android phones and iPhones, and inside Google Search. Every few months Google swaps in a smarter "brain" behind it. The newest one is called Gemini 4 Carbon, and it became widely available on November 10, 20261,6.
So will my Gemini get smarter?
Probably not. Carbon is only switched on for people who pay for Google AI Ultra, Google's top plan, which starts at $100 a month in the US5. Ultra subscribers can pick it from a menu in the Gemini app and in AI ModeJargon busterA chat-style section of Google Search where an AI writes the answer., the chat-style part of Google Search1,7. If you use Gemini for free, or pay for the cheaper Google AI Pro plan, nothing changes today. Pro subscribers keep the modelJargon busterThe AI "brain" behind a chatbot; companies swap in new models to make the assistant smarter. they already had, Gemini 4 Argon7.
Think of it like a restaurant adding a chef's tasting menu. It's impressive, but most diners are still ordering from the regular menu, and that menu didn't get worse.
What is it good at?
Google says Carbon is built for long, complicated jobs: writing and fixing big chunks of software, scientific research, and finding security holes in computer code2,5. Google itself still suggests the older Argon model for most everyday coding and knowledge work7.
Is it better than ChatGPT or Claude?
It's close at the top. An independent testing company, Artificial Analysis, first ranked Carbon number one on its scoreboard, then updated its tests. It now puts Carbon second, behind Anthropic's Claude Fable 5.2 and just ahead of Claude Opus 5.5 and OpenAI's GPT-6 Astra8,10. The scores are close enough that for chatting, writing emails or planning a trip, you likely wouldn't feel a difference.
What should I do?
Nothing, unless you're already an Ultra subscriber. Then look for "4 Carbon" in the model menu and save it for your hardest questions7.
Gemini 4 Carbon goes GA: Ultra-only for consumers, cheap-ish for developers
Two weeks after a gated debut, Google opened Gemini 4 Carbon to paying developers and AI Ultra subscribers. It's near the top of independent rankings and priced below Anthropic's flagships.
Google announced Gemini 4 Carbon on October 28 as its "next frontier modelJargon busterThe AI "brain" behind a chatbot; companies swap in new models to make the assistant smarter.," initially giving it only to cybersecurity defenders in its gated Fairwind ProgramJargon busterGoogle's gated program that gives approved cybersecurity defenders early or less-restricted model access.2,14. On November 10 it went generally available1,13. It's the second Gemini 4 model, following Gemini 4 Argon, which reached the Gemini app in October for Pro and Ultra subscribers12.
Who gets it
- Consumers: Google AI Ultra only ($100 a month and up in the US) in the Gemini app, AI ModeJargon busterA chat-style section of Google Search where an AI writes the answer. in Search and the AntigravityJargon busterGoogle's AI coding tool, where AI agents help write and fix software. coding tool1,5. In AI Mode it's for subscribers 18 and older in supported countries, with daily usage limits6.
- Pro subscribers: stay on Argon, with no announced timeline for Carbon7. The app's model picker now labels Argon "Code & analysis" and Carbon "Most complex work," and Gemini 3.1 Pro is gone7.
- Developers and businesses: the paid Gemini APIJargon busterA way for programmers to plug an AI model into their own apps and pay per use. in Google AI Studio and Gemini for business1.
- Security teams: a "Gemini 4 Carbon for cyber" variant with fewer guardrails stays behind a Fairwind allowlist1,5.
Not everyone is happy with the tiering. A thread on Google's own developer forum asks Google to bring Carbon to AI Pro15.
How it stacks up
Google's launch table shows Carbon with the top score in 15 of 17 rows. Those are Google's own numbers, and VentureBeat noted that OpenAI kept the lead on one software-engineering test and Anthropic on one computer-use test9.
Independent tests are more mixed. Artificial Analysis debuted Carbon at first place on its Intelligence Index v5.0 with 59 points, ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57)10. After the firm regraded one component, VentureBeat reports Claude Fable 5.2 at 65, Carbon at 60, Opus 5.5 at 59 and GPT-6 Astra at 578. On Artificial Analysis's coding-agent index, Carbon trails Claude Code running Opus 5.5, 59 to 6210. One warning sign: Carbon's hallucinationJargon busterWhen an AI states something false with confidence. rate on the AA-Omniscience test is 11%, versus 4% for Argon11.
The price play
Carbon launches at Argon's introductory API price, $2 per million input tokens and $10 per million output tokens, until November 30. From then on it's $4 and $201. VentureBeat calculates that's about a quarter of Fable 5.2's standard rates now, and half after the increase8. Artificial Analysis puts the cost of running its full test suite at $7.76 per task for Carbon versus $10.04 for Opus 5.510.
What to watch: whether Carbon reaches Pro subscribers, and whether developers stick around once the price doubles in December.
Gemini 4 Carbon: 1M-token shared context, text-out multimodal, $2/$10 rising to $4/$20
A long-horizon agenticJargon busterDescribes AI that carries out multi-step tasks on its own, like running code and checking results. modelJargon busterThe AI "brain" behind a chatbot; companies swap in new models to make the assistant smarter. built on Gemini 4 Argon, with a 1,048,576-token shared input/output budget, March 2026 knowledge cutoffJargon busterThe date after which the model has no built-in knowledge of the world. and big self-reported gains on agentic and security benchmarks.
Google's modelJargon busterThe AI "brain" behind a chatbot; companies swap in new models to make the assistant smarter. card describes Gemini 4 Carbon as built on Gemini 4 Argon and aimed at long-horizon agenticJargon busterDescribes AI that carries out multi-step tasks on its own, like running code and checking results. work across knowledge work, engineering, cybersecurity and scientific reasoning3. It was announced October 28 and went GA on the APIJargon busterA way for programmers to plug an AI model into their own apps and pay per use. and Google Cloud's Gemini Enterprise Agent Platform on November 10, alongside a cyber-permissive variant2,13.
Specs
| Spec | Value |
|---|---|
| Context windowJargon busterHow much text, images or audio the model can consider at once, measured in tokens. | 1,048,576 tokens, shared by input and output4 |
| Default output per request | 65,536 streaming / 32,768 non-streaming4 |
| Max output (long decode) | Up to the full window minus input length4 |
| Input modalities | Text, image, audio, video3,4 |
| Output | Text3,4 |
| Knowledge cutoffJargon busterThe date after which the model has no built-in knowledge of the world. | March 20263 |
| Caching | Implicit only, for shared prefixes of 4,096+ tokens; no explicit caching4 |
| API price, through Nov 30 | $2 in / $10 out per 1M tokens1 |
| API price, after Nov 30 | $4 in / $20 out per 1M tokens1 |
Note that the output budget counts thinking tokens and function calls, not just the visible answer4. Pricing on Google Cloud's pricing page includes extra tiers we couldn't fully reconcile, so check it for your deployment.
Google's self-reported benchmarks
These come from Google's launch post and aren't independently reproduced2,9:
| BenchmarkJargon busterA standardized test used to compare AI models; companies often run and report their own. | Argon | Carbon |
|---|---|---|
| Terminal-Bench 4.0 (AA) | 57.1% | 65.2%9 |
| Agents' Last Exam v1.1 | 69.0% | 77.4%9 |
| SWE-sweep | 4.3% | 9.6%9 |
| SEC-bench Pro (find vulns) | n/a | 88.2%9 |
| ExploitGym (exploit) | n/a | 54.8%9 |
| CWE-bench (patch) | n/a | 71.0%9 |
On FrontierSWE 2, Carbon's 61.2% trails GPT-6 Astra (65.5%) and Claude Opus (62.3%) in the same table9.
Independent results (Artificial Analysis)
- Intelligence Index v5.0: 59 at launch10; 60 at high reasoning after a regrade, behind Claude Fable 5.2 at max effort (65), ahead of Opus 5.5 max (59) and GPT-6 Astra max (57)8.
- Coding Agent Index: 59, versus 62 for Claude Code with Opus 5.5 at max10.
- AA-Omniscience v1.1: 54, ahead of Opus 5.5 (48) and Argon (46), but with an 11% hallucinationJargon busterWhen an AI states something false with confidence. rate versus Argon's 4%11. The rate is incorrect answers divided by incorrect, partial and abstained answers, not all answers11.
- Cost to run the index: $7.76 per task at high, versus $10.04 for Opus 5.5 max10 and $26.54 for Fable 5.2 max8.
Caveats and open questions
Scores move when evaluators update their suites, as Artificial Analysis's own regrade shows8. The price per tokenJargon busterA small chunk of text, roughly three-quarters of a word, used to measure and bill AI usage. is low, but Carbon spends many tokens on reasoning, so cost per task matters more than list price. The cyber variant's guardrail differences aren't publicly documented beyond it being gated behind Fairwind1.
Quick quiz: did it stick?
Q1 What do you need to use Gemini 4 Carbon in the Gemini app?
Q2 When did Google make it generally available?
Q3 What does an independent tester say about it?
By the numbers
How we got here
- Oct 13, 2026Gemini 4 Argon reaches the Gemini app for Pro and Ultra subscribers12
- Oct 28, 2026Google announces Gemini 4 Carbon, limited to Fairwind cyber defenders2,14
- Nov 10, 2026Carbon goes generally available to paying developers and Ultra subscribers1,6
- Nov 30, 2026 · coming upIntroductory API price ends1
- Dec 1, 2026 · coming upAPI price rises to $4 / $20 per 1M tokens1,6
What should you do?
Jargon buster
- Model
- The AI "brain" behind a chatbot; companies swap in new models to make the assistant smarter.
- Google AI Ultra / Pro
- Google's paid AI subscription plans; Ultra is the most expensive tier.
- AI Mode
- A chat-style section of Google Search where an AI writes the answer.
- Antigravity
- Google's AI coding tool, where AI agents help write and fix software.
- API
- A way for programmers to plug an AI model into their own apps and pay per use.
- Token
- A small chunk of text, roughly three-quarters of a word, used to measure and bill AI usage.
- Context window
- How much text, images or audio the model can consider at once, measured in tokens.
- Benchmark
- A standardized test used to compare AI models; companies often run and report their own.
- Hallucination
- When an AI states something false with confidence.
- Fairwind Program
- Google's gated program that gives approved cybersecurity defenders early or less-restricted model access.
- Agentic
- Describes AI that carries out multi-step tasks on its own, like running code and checking results.
- Knowledge cutoff
- The date after which the model has no built-in knowledge of the world.
Sources22 pages · Google, Google DeepMind, Google Cloud and moreShowHide
The small numbers in the text point to these pages. Everything is written in our own words. How we work · Spotted a mistake? Tell us
- 1Gemini 4 Carbon is now available · Google
- 2Gemini 4 Carbon: our next frontier model · Google
- 3Gemini 4 Carbon model card · Google DeepMind
- 4Gemini 4 Carbon (Gemini Enterprise Agent Platform docs) · Google Cloud
- 5Google's Gemini 4 Carbon is now widely available around the world · Engadget
- 6Google rolls out Gemini 4 Carbon to AI Ultra subscribers · Android Authority
- 7Gemini 4 Carbon rolling out to Gemini app for AI Ultra, Argon remains · 9to5Google
- 8Google Gemini 4 Carbon is now publicly available, and it undercuts Claude Fable 5.2 and Opus 5.5 on API price · VentureBeat
- 9Google unveils Gemini 4 Carbon, retaking the world's no. 1 spot on updated, harder benchmarks, but access is limited · VentureBeat
- 10Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5 · Artificial Analysis
- 11AA-Omniscience v1.1: Knowledge and Hallucination Benchmark · Artificial Analysis
- 12Google's Gemini 4 Argon is now available in the Gemini app and AI Mode in Search · Engadget
- 13Gemini Enterprise Agent Platform release notes · Google Cloud
- 14Google announces Gemini 4 Carbon with 'significant gains across coding and agentic tasks' · 9to5Google
- 15Unfair Tier Gating: Include Gemini Carbon Access in Google AI Pro · Google AI Developers Forum
- 16Google AI Pro & Ultra plans · Google
- 17Google AI plans (United Kingdom) · Google
- 18Google AI plans (France) · Google
- 19Google AI plans (Australia) · Google
- 20Google AI plans (Canada) · Google
- 21Google AI plans (India) · Google
- 22Google AI Ultra が月額 14,500 円から利用可能に。上位プランも 32,000 円に値下げ · HelenTech
Read next
AI
ChatGPT Wallet lets OpenAI's AI agents buy things for you, with a Face ID check each time
OpenAI gave its AI helpers a way to pay for things with your credit card. You still say yes to every purchase with your face or fingerprint, and only a small group of paying US users can try it.
ChatGPT Wallet lets OpenAI's always-on dots agents complete checkouts on ordinary store websites using network tokens. It's narrow at launch, but it puts OpenAI in the card-network-backed agent payments race with Google, Visa and Mastercard.
Wallet swaps the PAN for a network token scoped to one purchase, gates each charge behind device biometrics, and rides existing checkout forms. Merchants link tokens to accounts via the Payment Account Reference.
Wearables
Samsung and Google's first AI glasses go on preorder from $395, shipping December 10
Samsung and Google made a pair of normal-looking glasses with a camera, tiny speakers and an AI assistant you talk to. They start at $395 and arrive December 10, in the US and South Korea only.
Samsung's Intelligent Eyewear opened for preorder November 12 at $395 to $489, shipping December 10. It's display-free, camera-equipped and runs Gemini, like Meta's Ray-Bans, but undercuts their $449 Gen 3.
A display-free Android XR device on Qualcomm's Snapdragon AR1 Gen 1 with 2GB RAM, 64GB storage and a 12MP camera, offloading Gemini to the paired phone. Wi-Fi 6E, Bluetooth 5.4, nine hours claimed.
Computers
Googlebooks get their first big fix: up to 6GB of RAM back and about 30 bugs squashed
Googlebooks, Google's new laptops that run on Android, are getting their first big update. It fixes crashes and frees up memory so they feel smoother. Regular Chromebooks aren't affected.
Five weeks after launch, Google's November Googlebook OS update cuts on-device AI memory use by 3GB to 6GB on the Intel models and fixes around 30 issues. Chromebook migration details are still missing.
The November build trims memory held by the on-device AI stack on the three Panther Lake models, adds browser reclaim tuning on all five, and ships the 2026-11-01 security patch level.