This guide is built from Google’s model documentation and pricing pages as summarised by independent trackers, plus third-party benchmarks. I have not run Gemini 4 Carbon on my own workloads, and every benchmark below is attributed, with Google’s own claims labelled as such.
Gemini 4 Carbon is Google’s most capable model, generally available since 10 November 2026, and for the next twelve days it costs the same as its cheaper sibling. Until 30 November, Carbon is priced at $2 per million input tokens and $10 output. On 1 December, every rate doubles, to $4 and $20.
That pricing clock is the most useful thing to know. Carbon is a genuinely strong model, second on SemiAnalysis’s demanding GDPx agent benchmark and near the top of the main independent index. But it thinks hard and spends tokens freely, so its real cost per task depends heavily on how you configure it. Test it this month, while testing is cheap.
Table of Contents
- Fact 1: Who Can Use Gemini 4 Carbon
- Fact 2: Gemini 4 Carbon Pricing Doubles on 1 December
- Fact 3: Specs and Limits
- Fact 4: Gemini 4 Carbon vs Argon
- Fact 5: The Benchmarks, Google’s and Everyone Else’s
- Fact 6: Security Is Google’s Headline Claim
- Fact 7: How Gemini 4 Carbon Compares on Price
- How to Test Gemini 4 Carbon Before 1 December
- Five Mistakes to Avoid When You Switch
- Who Should Use Gemini 4 Carbon
- FAQ About Gemini 4 Carbon
- Verdict
- Sources and Further Reading

Fact 1: Who Can Use Gemini 4 Carbon
Google announced general availability in a blog post dated 10 November. Carbon had launched about two weeks earlier, but only for cybersecurity defenders through Google’s Fairwind Program, which is why some articles from late October say it is not publicly available. That is no longer true.
Here is where it runs now, per Google’s pages as compiled by BrainDetox:
- The API, through Google AI Studio and Gemini for business, for paying developers and enterprises. The model ID is
gemini-4-carbon. - The Gemini app and AI Mode in Search, rolling out to Google AI Ultra subscribers, who can also try it in Google Antigravity and AI Studio.
- Scientists can apply for Antigravity’s highest usage tier free for six months, subject to eligibility and capacity.
One important exclusion: Gemini 4 Carbon is not available on the free tier. If you have been testing Gemini on free credits, you will need a paid account to try this one.
Where you may meet it without writing code
Not everyone will reach the model through the API. Google is putting it into its own products for paying subscribers first, starting with the Gemini app and AI Mode in Search for Ultra members. That matters for marketers and publishers as much as developers, because the model answering questions in AI Mode shapes which sources get cited and how products are described.
It also appears in Google Antigravity, the company’s agentic development environment, where Ultra subscribers can try it on coding tasks. If your team already uses Antigravity, that is the lowest-friction way to compare it with whatever model you use today, without touching billing settings.
Finally, the Fairwind Program remains a separate route for security teams. Defenders who joined before general availability keep access to a version Google describes as tuned for cyber work, which is worth asking about if you run a security operations team.
Fact 2: Gemini 4 Carbon Pricing Doubles on 1 December
Here is the full rate card, per million tokens, through 30 November and from 1 December 2026.
| Mode | Input | Output (including thinking) | Cached input |
|---|---|---|---|
| Standard | $2 then $4 | $10 then $20 | $0.10 then $0.20 |
| Batch | $1 then $2 | $5 then $10 | $0.05 then $0.10 |
| Flex | $1 then $2 | $5 then $10 | $0.05 then $0.10 |
| Priority | $3.60 then $7.20 | $18 then $36 | $0.18 then $0.36 |
Every paid account moves to the new price on the same day, so there is no grandfathering for early users. Implicit caching is on by default for prompts of at least 4,096 tokens, with a 95% discount on cache hits. Google Search grounding includes 5,000 free requests a month, then costs $14 per 1,000, with a daily cap of 1,500.
Note that output pricing includes thinking tokens. Gemini 4 Carbon thinks a lot, so the output line is where most of your bill will come from. Paid-tier data is not used to improve Google’s products, according to the pricing page.
Fact 3: Specs and Limits
The context window is 1,048,576 tokens, shared between input, thinking and output. Maximum output is 65,536 tokens for synchronous calls, or up to the full million with long decode in background mode.
Carbon accepts text, image, video, audio and PDF input and returns text. It has three thinking levels, low, medium and high, with medium as the default. It supports code execution, computer use, file search, function calling, Google Maps and Search grounding, structured outputs, URL context and implicit caching.
What it does not support matters too: no audio generation, no image generation, no Live API and no explicit caching. If you rely on any of those with another Gemini model, keep that model in your stack.
Fact 4: Gemini 4 Carbon vs Argon
Gemini 4 Argon, announced on 30 September and generally available from 13 October, is the model most developers will compare Carbon against. They share a lot.
| Gemini 4 Carbon | Gemini 4 Argon | |
|---|---|---|
| Price per 1M tokens | $2 / $10, then $4 / $20 | $2 / $10, then $4 / $20 |
| Context | 1M tokens | 1M tokens |
| Thinking per turn | More tokens, higher latency | Fewer tokens, faster |
| Best for | Root-cause debugging, architecture, proofs, security audits | Day-to-day engineering and knowledge work |
| GDPx score | 10.9% | 7.3% |
| Free tier | No | Check current terms |
The switch is easy. Request schemas need no changes; only the model string changes. Google’s own docs advise evaluating at all three thinking levels and choosing the lowest one that meets your quality bar, which is exactly the right instinct.
My rule of thumb: start every new task on Argon. Move to Gemini 4 Carbon only for the specific steps where Argon’s answers are not good enough. Since both cost the same per token, the difference is all in how many tokens Carbon burns thinking.
Fact 5: The Benchmarks, Google’s and Everyone Else’s
Google’s claims are strong. It says Gemini 4 Carbon ranks first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, at more than twice lower cost per task than the next-best external model, which it does not name. It places second on the Multimodal Knowledge Agent benchmark at more than six times lower average cost per task than the leader. In finance, Google says it matched the best model on Hebbia’s agentic evaluation and topped 23 configurations on Balyasny’s benchmark before release.
Independent numbers are more useful. On SemiAnalysis’s GDPx benchmark, which tests 206 long professional tasks, Gemini 4 Carbon scored 10.9%, second only to Claude Opus 5.5 at 12.0% and ahead of GPT-6 Astra at 9.6%. That is a genuinely strong result on one of the hardest agent tests published.
On the Artificial Analysis Intelligence Index v5, release trackers report Carbon at around 59 to 60, behind Claude Fable 5.2 at 65 and ahead of Kimi K3.1 Pro at 55. The exact figure varies between snapshots, so check the live page before quoting it.
Fact 6: Security Is Google’s Headline Claim
Carbon launched first to cyber defenders for a reason. Google publishes a long list of security figures, and they are worth knowing even if they are self-reported.
- Blocked 99% of attack trajectories from an adaptive red-teaming agent in simulated environments.
- A per-attempt breach rate of 0.8%, down from 3.1% on Argon.
- An automated cyberattack success rate of 2.17% in Gray Swan’s testing, against 9.86% for Argon.
- Zero new universal jailbreaks found by Trajectory Labs across 132 hours and more than 13,000 requests.
- 92% on a 375-task honesty benchmark, against 70% for Argon.
Google is also candid that resistance to prompt injection remains an area for improvement. If you are building agents that read untrusted content, that admission matters more than any of the headline numbers. Our guide to prompt injection and agent security explains why no model alone solves this.
Fact 7: How Gemini 4 Carbon Compares on Price
At the introductory rate, Gemini 4 Carbon is one of the cheapest frontier-class models per token. After 1 December it moves into the middle of the pack.
| Model | Input per 1M | Output per 1M |
|---|---|---|
| Gemini 4 Carbon (to 30 Nov) | $2 | $10 |
| Gemini 4 Carbon (from 1 Dec) | $4 | $20 |
| GPT-6.1 Sol | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Fable 5.2 | $8 | $40 |
| GPT-6 Astra | $10 | $50 |
Per-token price is only half the story. Eesel’s analysis of Argon on the Artificial Analysis index found it scored 53 at high effort for about $1.99 per task at the introductory price, rising to about $3.98 at the regular price, against Claude Opus 5.5 at high effort scoring 54 for about $1.82. Carbon thinks more than Argon, so expect its cost per task to be higher still. Our guide to LLM API cost optimisation covers how to keep thinking-heavy models under control.
How to Test Gemini 4 Carbon Before 1 December
You have until the end of November to evaluate at half price. Here is how I would spend that time.
Pick twenty real tasks. Take them from your own logs, weighted towards the hard ones where your current model struggles. Easy tasks will not show what Carbon adds.
Run each at low, medium and high thinking. Record quality, latency and total tokens, including thinking. Many tasks will be fine at low, which roughly halves the cost compared with the default.
Run the same tasks on Argon. Since the per-token price is identical, any quality gain from Carbon has to justify the extra thinking tokens.
Multiply by two. Whatever your test costs this month, double it to get the December price. Budget on that number, not the introductory one.
Use batch for anything not urgent. Batch and Flex are half the standard rate in both periods. For overnight jobs, that alone offsets the December increase.
Five Mistakes to Avoid When You Switch
Switching models looks like a one-line change. In practice, teams that move fast tend to trip over the same handful of problems.
Budgeting on the introductory price. This is the obvious one, and it still catches people. A cost model built on November numbers will be wrong by a factor of two from 1 December. Put the regular rate in every spreadsheet from the start.
Leaving thinking on the default. Medium is the default level. For classification, extraction and short answers, low is often enough, and the token saving compounds across every call. Test each task type separately rather than picking one level for everything.
Setting the output limit too low. Google suggests an output limit of at least 32,768 tokens because thinking shares the output budget. A tight limit can cut answers off mid-thought, which looks like a quality problem but is really a configuration one. For generations beyond 64K, use background mode.
Forgetting the missing features. No image or audio generation, no Live API, no explicit caching. If a workflow quietly depends on one of those, it will break when you swap the model string. List your dependencies before you migrate.
Skipping a regression run. A stronger model is not always better at your particular prompt. Keep twenty or thirty reference tasks with known good answers, run them before and after any switch, and compare. It takes an hour and prevents most surprises. Our roundup of LLM evaluation tools covers frameworks that automate this.
None of these is specific to Google. They apply to any thinking model with tiered pricing, which by now means almost every frontier model on the market.
Who Should Use Gemini 4 Carbon
| If you are | Start with | Why | Link |
|---|---|---|---|
| Running long, hard agent tasks | Gemini 4 Carbon at low or medium thinking | Second on GDPx, ahead of GPT-6 Astra | Gemini API pricing |
| Doing everyday coding and writing | Gemini 4 Argon | Same price per token, faster and fewer thinking tokens | Google AI for Developers |
| Security-sensitive agent work | Carbon, with your own guardrails | Strong reported red-team results, injection still a weak spot | Agent security guide |
| Cost-focused batch processing | Carbon or Argon on Batch or Flex | Half the standard rate in both pricing periods | Cut LLM API costs |
| On the free tier | Argon or older Gemini models | Carbon is paid-only | Google AI Studio |
Related on this site: for more context, see Google Antigravity Six Months After 2.0 and Siri AI Two Months In.
FAQ About Gemini 4 Carbon
Is Gemini 4 Carbon available to everyone?
It is generally available to paying developers and enterprises through the Gemini API since 10 November 2026, and rolling out to Google AI Ultra subscribers in the Gemini app. It is not on the free tier.
How much does Gemini 4 Carbon cost?
$2 per million input tokens and $10 output until 30 November 2026, then $4 and $20 from 1 December. Batch and Flex are half those rates.
What is the difference between Gemini 4 Carbon and Argon?
They share pricing, context and tools. Carbon spends more tokens thinking for harder reasoning tasks, while Argon is faster and cheaper per turn for everyday work.
What is the Gemini 4 Carbon context window?
1,048,576 tokens shared across input, thinking and output, with up to 65,536 output tokens synchronously or the full million in background mode.
How does Gemini 4 Carbon score on benchmarks?
It placed second on SemiAnalysis’s GDPx at 10.9%, and trackers report around 59 to 60 on the Artificial Analysis Intelligence Index v5. Google claims first place on several partner benchmarks.
Can Gemini 4 Carbon generate images or audio?
No. It outputs text only and does not support audio generation, image generation or the Live API.
Verdict
Test Gemini 4 Carbon before 1 December. It is one of the strongest agent models available, second only to Opus 5.5 on GDPx, and for the rest of November it costs the same as the cheaper Argon. Use that window to find out which of your tasks genuinely need it.
The caveat is that Carbon is a heavy thinker, and output pricing includes those thinking tokens. Once rates double, a model that burns twice the tokens at twice the price can cost four times what you expected. Most of the benchmark wins are Google’s own numbers, and Google itself admits prompt injection is still a weak spot.
Your next step: set up a paid Gemini API key today, run twenty hard tasks at low and medium thinking on both Carbon and Argon, and budget for December at double what you spend this week.
Sources and Further Reading
- BrainDetox on Carbon’s general availability, for access, limits, the full pricing table and Google’s security figures
- BuildMVPFast on the December price change, for how the introductory rates expire for every account
- Eesel’s pricing analysis, for cost per task comparisons against Opus and Astra
- Tech Insider on Argon’s launch, for the Argon timeline and the pricing cliff
- Let’s Data Science on GDPx, for the independent agent benchmark results
More model coverage on this site: GPT-6 Astra, Claude Fable 5.1 and AI models compared. If you use Google’s creative tools, see our guide to Google Flow agent prompts.


[…] Related on this site: for more context, see Claude Code vs Codex vs Cursor, Copilot AI Credits Explained and Gemini 4 Carbon. […]