What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind's frontier model. Google announced it on 28 October 2026 and at first gave it only to vetted cyber defenders in its invite-only Fairwind programme. On 10 November 2026 it opened it to developers, businesses and consumers: the paid tier of the Gemini API in AI Studio and Android Studio, Gemini for business, Google Antigravity and CodeMender, and the Gemini app and AI Mode in Search for Google AI Ultra subscribers. Google AI Pro does not include it. Fairwind continues for a separate cyber-permissive version.
The model ID is gemini-4-carbon, listed as stable. It reads text, images, video, audio and PDF files and writes text. Input and output share a window of 1,048,576 tokens; one response is capped at 65,536 tokens by default, and a long-decode mode stretches output to the full million minus the input. Thinking has low, medium and high levels, with medium the default. Google's knowledge cutoff is March 2026.
API pricing matches Argon: $2 per million input tokens and $10 per million output tokens, thinking included, until 30 November 2026, then $4 and $20. Batch and Flex run at half price, Priority costs 80% more, and only implicit caching is supported, at $0.10 per million cached input tokens (then $0.20). There is no free tier.
How much does Gemini 4 Carbon cost?
| Plan | Price | What you get |
|---|---|---|
| Until 30 Nov 2026 | $2 in / $10 out per 1M tokens | Cached input $0.10; Batch and Flex half price |
| From 1 Dec 2026 | $4 in / $20 out per 1M tokens | Cached input $0.20; Batch and Flex half price |
| Google AI Ultra | $99.99/month (5x) or $199.99/month (20x) | Carbon in the Gemini app, AI Mode and Antigravity, in selected countries; Google AI Pro gets Argon only |
API: $2 input / $10 output per 1M tokens until 30 Nov 2026, then $4 / $20; cached input $0.10 (then $0.20); Batch and Flex half price; Priority $3.60 / $18 (then $7.20 / $36). No free tier. Gemini app: Google AI Ultra, $99.99/month (5x) or $199.99/month (20x).
Compare this price with other models
Is Gemini 4 Carbon worth it?
Carbon opened to the public on 10 November 2026, 13 days after a launch limited to Google's Fairwind cyber programme. Artificial Analysis now scores it 60 on Intelligence Index v5.0.1 at high thinking, fourth of 74 models. It sits behind Claude Fable 5.2 (65 at max effort) and ahead of Claude Opus 5.5 (59), GPT-6 Astra (57) and Argon. At launch Vals put it second on its own index at 70.36%, within its ±0.97 margin of Claude Sonnet 5.5 (70.85%) and Opus 5.5 (70.18%), and first on the Vals Finance Index and Legal Research Bench.
Google's launch table is self-reported and has no error margins. It claims the top score on 15 of 17 rows, including 96.2% on GraphWalks long context against Astra's 71.8% and Opus's 66.8%. The cyber figures, 88.2% on SEC-bench Pro and 54.8% on ExploitGym, were run without guardrails, and the model card says they reflect the Fairwind version, not the public one. Even on Google's table it trails on FrontierSWE 2 (61.2% against Astra's 65.5% and Opus's 62.3%) and on OSWorld computer use (81.5% against Opus's 87.2%). VentureBeat found it ahead of Fable 5.2 on Terminal-Bench 4.0 (65.2% against 61.6%) but behind on FrontierCode 2.0 (40.6% against 49.1%).
The public model has tighter cyber safeguards than Argon. Google says they cut the per-attempt breach rate of its red-teaming agent to 0.8%, against 3.1% on Argon, and Gray Swan's automated cyberattack success rate to 2.17%, against 9.86%. Defenders who need exploit validation or malware analysis apply to Fairwind for gemini-4-carbon-cyber-permissive. Reliability data is mixed: per VentureBeat, at launch its hallucination rate on Artificial Analysis's Omniscience test rose from Argon's 4% to 11%, still far below the 50% recorded for Opus 5.5 and Astra.
It is slow and wordy. Artificial Analysis measures 65.84 seconds to the first answer token against a median of 13.45, 77 tokens a second, and 205 million output tokens to run its index against a median of 72.8 million. At the $4/$20 rate due on 1 December that is $7.76 per index task: below Opus 5.5's $10.04 and far below Fable 5.2's $26.54 at max effort, above Astra's $6.25.
That price is the case for it. From December Carbon costs the same per token as Opus 5.5 and a score one point higher, and until 30 November it costs half. Teams that need Fable 5.2's lead pay four times Carbon's December per-token rate for it. The catches are the wait before the first token, no explicit caching, and, in the Gemini app, a $99.99 Ultra plan.
Where it earns its place
A platform lead at a company already running Gemini 4 Argon in production
Move hard agent tasks to Carbon without blowing the budget
Change the model ID to gemini-4-carbon; per-token prices are identical, so the bill moves with token use. Carbon spent 205 million output tokens on the Artificial Analysis index, and costs $7.76 per index task against Argon's $7.68 at the December rate. A job of 10M uncached input and 1M output tokens costs $30 until 30 November and $60 after.
A head of AI at a bank or law firm comparing frontier models
Pick a model for financial analysis and legal research agents
At launch Carbon led the Vals Finance Index at 66.35% and Legal Research Bench at 55.77%, and from December it costs the same per token as Opus 5.5. Fable 5.2 scores higher on Artificial Analysis but costs twice as much per token. Paid-tier data is not used to improve Google's products.
A national CERT or critical-infrastructure security team
Find, validate and patch vulnerabilities in large codebases before attackers do
The public model refuses more offensive work than Argon, so apply to Fairwind for gemini-4-carbon-cyber-permissive, which Google offers in the global and us regions with high thinking by default. Google's 88.2% on SEC-bench Pro and 71.0% on CWE-bench are self-reported, run without guardrails and reflect Fairwind access.
Before you commit
- Setup
- Model ID gemini-4-carbon on the Gemini API paid tier; a billing-enabled project is required because there is no free tier. Thinking level is low, medium (default) or high; setting thinking_budget or MINIMAL returns a 400 error, and temperature, top_p and top_k are ignored. Moving from gemini-3.1-pro-preview means removing explicit caching and prefilled model turns.
- Integrations
- Gemini API in AI Studio and Android Studio with Batch, Flex and Priority; function calling, Google Search and Maps grounding (5,000 free requests a month shared across Gemini 3 and later, then $14 per 1,000), code execution, computer use, file search, URL context and structured outputs. No Live API. Google Cloud's Gemini Enterprise Agent Platform lists it in the global, us and eu regions. Also in Gemini for business, Antigravity, CodeMender, and the Gemini app and AI Mode for AI Ultra subscribers in selected countries.
- Data and privacy
- On the paid API tier Google says data is not used to improve its products, but prompts and responses are logged for 55 days for abuse monitoring. Google Cloud documents a store=false mode for long decodes in zero-data-retention setups. The public model has strengthened cyber and CBRN misuse safeguards, system-level filtering and usage monitoring; the cyber-permissive version is limited to Fairwind-approved projects.
- Cost at scale
- 10,000 document reviews a month at 200K tokens in and 4K out cost about $4,400 until 30 November and $8,800 after; Batch or Flex halves either figure. Thinking bills as output, and Artificial Analysis counts $7.76 per index task at the $4/$20 rate. Implicit caching applies to shared prompt openings of 4,096 tokens or more.
- Limits
- 1,048,576 tokens shared between input, thinking and output; 65,536 output tokens per streaming response and 32,768 non-streaming by default, with long decode (v1beta1 only) up to the full window. Per-model RPM and TPM are shown only in AI Studio; batch queues hold 5M tokens at Tier 1, 500M at Tier 2 and 1B at Tier 3. Google lists occasional slowness or timeouts as a known limitation.
What’s good
- 60 on Artificial Analysis Intelligence Index v5.0.1, ahead of Opus 5.5 and GPT-6 Astra
- Same per-token price as Argon and, from December, Opus 5.5; half that until 30 November
- $7.76 per index task against $10.04 for Opus 5.5 and $26.54 for Fable 5.2 at max
- Native video, audio, image and PDF input in one 1M-token window; Batch and Flex at half price
What’s not
- Behind Claude Fable 5.2 on Artificial Analysis (60 against 65)
- 66-second median wait for the first answer token, and very verbose
- No free tier, no explicit caching; Gemini app access needs AI Ultra at $99.99/month
- Trails Astra and Opus 5.5 on FrontierSWE 2, and Opus on OSWorld
Best for
Gemini API and Google Cloud teams that want frontier reasoning at Opus 5.5's price or below, especially on finance, legal and long-document work.
Gemini 4 Carbon alternatives
Gemini 4 Carbon: common questions
- Is Gemini 4 Carbon worth it?
- Carbon is now the best model anyone can buy from Google, and at Argon's per-token price it beats Opus 5.5 on Artificial Analysis for the same money; Claude Fable 5.2 is still ahead. We rate it 4.6 out of 5.
- How much does Gemini 4 Carbon cost?
- API: $2 input / $10 output per 1M tokens until 30 Nov 2026, then $4 / $20; cached input $0.10 (then $0.20); Batch and Flex half price; Priority $3.60 / $18 (then $7.20 / $36). No free tier. Gemini app: Google AI Ultra, $99.99/month (5x) or $199.99/month (20x).
- What are the downsides of Gemini 4 Carbon?
- Behind Claude Fable 5.2 on Artificial Analysis (60 against 65); 66-second median wait for the first answer token, and very verbose; no free tier, no explicit caching; Gemini app access needs AI Ultra at $99.99/month; trails Astra and Opus 5.5 on FrontierSWE 2, and Opus on OSWorld.
- Who is Gemini 4 Carbon best for?
- Gemini API and Google Cloud teams that want frontier reasoning at Opus 5.5's price or below, especially on finance, legal and long-document work.
- How much does Gemini 4 Carbon cost for a team or in production?
- 10,000 document reviews a month at 200K tokens in and 4K out cost about $4,400 until 30 November and $8,800 after; Batch or Flex halves either figure. Thinking bills as output, and Artificial Analysis counts $7.76 per index task at the $4/$20 rate. Implicit caching applies to shared prompt openings of 4,096 tokens or more.
- Is Gemini 4 Carbon safe for company data?
- On the paid API tier Google says data is not used to improve its products, but prompts and responses are logged for 55 days for abuse monitoring. Google Cloud documents a store=false mode for long decodes in zero-data-retention setups. The public model has strengthened cyber and CBRN misuse safeguards, system-level filtering and usage monitoring; the cyber-permissive version is limited to Fairwind-approved projects.
- What are the best Gemini 4 Carbon alternatives?
- The closest alternatives we have reviewed are Claude Fable 5.2, Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon.
Answers checked Nov 11, 2026.
