Quick Answer: Gemini 4 Carbon is Google's strongest frontier model, now available through the API in Google AI Studio and Gemini for business at $2 per million input tokens and $10 per million output tokens. It targets the hardest tasks: multi-step debugging, long-horizon agentic workflows, cybersecurity, and open-ended scientific research. For most routine engineering and knowledge work, Google still recommends its sibling model, Gemini 4 Argon.
Most launch posts call a new model "smarter" and stop there. That tells you almost nothing about whether you should switch. What you need is the workload it's built for, what it costs per finished task rather than per token, and where the vendor numbers stop being useful. Google's announcement of Gemini 4 Carbon covers all three, and the details matter more than the headline.
Gemini 4 Argon vs Carbon: Pick by Task, Not by Hype
Google positions Carbon as the model for work that pushes against the limits of reasoning. Argon stays the default for most software engineering and knowledge work. The split is less about capability than about how many tokens each model spends.
| Dimension | Gemini 4 Argon | Gemini 4 Carbon |
|---|---|---|
| Intended workload | Majority of software engineering and knowledge work | Complex multi-step debugging, open-ended scientific research |
| Token behavior | Lighter reasoning per task | Uses more tokens to explore, reason, and run verification loops |
| Introductory API price | $2 input / $10 output per million tokens | $2 input / $10 output per million tokens |
| Rollout | Available now | API in Google AI Studio and Gemini for business for paying developers and enterprises; Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search |
The practical question is simple: does your task need the extra verification passes? A nightly report pipeline almost certainly doesn't. A root-cause investigation across a dozen microservices probably does.
Why Cost Per Task Matters More Than Price Per Token
Here's the counter-intuitive part. Carbon and Argon share the same introductory per-token price, so a naive comparison says they cost the same. That comparison is wrong, because Carbon generates more tokens per job. A model that charges the same rate but finishes in fewer retries can still come out cheaper overall.
Google claims exactly this pattern. Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench, at more than 2X lower cost per task than the next best external model. The mechanism is cause and effect: when a model verifies its own work before returning it, fewer bad outputs reach you, and fewer follow-up calls are needed to fix them.
That said, these are vendor-selected benchmarks, run pre-release. Don't take the 2X figure as a promise for your codebase. Before you migrate anything, run a fixed set of 50 to 100 real tasks through both models and track:
- Total tokens consumed per completed task, not per call
- Number of retries or human corrections needed
- Wall-clock time to a passing result
- Dollar cost per accepted output
If Carbon wins on that list for your workload, the per-token price stops mattering. If it doesn't, Argon is the cheaper choice even when Carbon's benchmarks look better.
Multi-Agent Work With /swarm in Google Antigravity
The most architecturally interesting part of the launch is the /swarm feature in Google Antigravity, currently in preview. It turns a single prompt into a coordinated team of agents you can track in real time. Google's suggested pattern is specific and worth copying:
- Assign Carbon to roles that need the deepest reasoning: architectural design, coordination, and auditing.
- Pair it with faster, lighter models for the parallel implementation work.
- Give each parallel worker an isolated workspace so they don't overwrite each other's changes.
- Add adversarial verifiers that test every change before it merges.
The failure mode to watch is coordination cost. Every agent-to-agent message consumes tokens, and a swarm that spends most of its budget talking to itself will be expensive without being better. Keep the number of agents small until you've measured where the loops actually add value. Most people add agents first and measure later, and that's backwards.
Document-Heavy Work: The MKA Benchmark and Real Results
Legal, finance, and enterprise operations are where Carbon's claims get concrete. Existing benchmarks mostly test text, but real analyst work involves charts, blueprints, schematics, and tables. To address that gap, Google partnered with Artificial Analysis to launch the Multimodal Knowledge Agent (MKA) benchmark. Carbon ranks #2 on MKA while costing more than 6X less on average per task than the top-performing model.
Some of the examples in the announcement show what this looks like under load. A team of Carbon agents, given seven public City Council planning PDFs (206 extracted engineering drawing views) and one prompt, built an interactive 906K-triangle Three.js reconstruction of a Google office across 10 overnight loops. The output is impressive, but the stronger signal is the process: the agents extracted, cross-checked, and modeled from dense source documents without hand-holding.
On financial benchmarks, Carbon matched the best tested model on Hebbia's E2E Financial Agentic evaluation and ranked at the top out of 23 frontier model configurations on Balyasny Asset Management's General Financial Intelligence benchmark. Treat these as directional. Vendor-chosen evaluations rarely match your document mix, and a 1-point gap on someone else's test tells you little about your own contracts or filings.
Cybersecurity Capability and the Fairwind Program
Carbon's strongest cyber results are gated. Google first released it to a set of cybersecurity defenders through the Fairwind Program, and the headline numbers apply to that access path. Carbon leads on SEC-bench Pro for finding vulnerabilities, ExploitGym for exploiting them, and CWE-bench v1 for patching them.
The internal numbers are striking. In Project Fortify, Google's autonomous scanning, triage, and remediation system, Carbon hit 100% recall on scanning and 95% precision on triage on Google's internal benchmark. On open-source repository testing, Google reports over a 25% relative improvement in recall over Argon.
External results are more interesting than the internal ones. As part of Wiz's Scan for Good program, the team used Carbon against two major SaaS providers and found 0-day vulnerabilities in under two hours. That's the kind of compression that changes how a security team plans a week. The caveat is dual use. The same capability that helps a defender find a flaw helps an attacker, which is why access is controlled and why disclosure timelines matter. Google says details of one open-source bug will follow after responsible disclosure is complete.
Scientific Reasoning: Real Gains With Honest Limits
Google's science claims fall into two groups: concrete deployments and early research results. Axiom Bio, a drug toxicity company, reports that Carbon is now their top-performing model for translational science, and it surfaced toxicity mechanisms they hadn't found before. Axiom is moving production analyses to both Carbon and Argon.
Google's internal work is more ambitious. Its AlphaProof Code framework used Carbon to generate millions of lines of Lean code aimed at formally proving security properties of the Signal protocol. Google says that work would historically take expert teams years. Its protein work is also notable: across seven low-homolog protein families from the CASP17 blind test set, Carbon increased retrieved homologs by 40-fold on average, without retraining the downstream structure models.
Here's the catch. A 40-fold increase in retrieved homologs is a retrieval result, not a structure prediction result. Google says broader benchmarking is still underway. Keep the claim in its proper box: better search upstream, with the downstream accuracy gains yet to be proven.
Access Today and What It Costs
For most readers, the practical question is who can use Carbon now and what it costs. Based on the announcement:
- API access: Paying developers and enterprises can use Carbon in Google AI Studio and Gemini for business at the introductory price of $2 per million input tokens and $10 per million output tokens, the same as Argon.
- Consumer and Ultra access: Google AI Ultra subscribers get Carbon in Google Antigravity, the Gemini app, and AI Mode in Search.
- Scientists program: Google opened applications for 10,000 scientists to access Argon in Google Antigravity free for six months. Carbon access is now being expanded to current and future participants, and the program will carry future frontier models. Google reports thousands of scientists already onboarded, with applications still open.
Note that "introductory" pricing can change. Build your cost model on current rates and recheck before committing a production budget.
Frequently Asked Questions
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google's strongest frontier model, released on November 10, 2026. It's built for complex, multi-step work such as deep debugging, long-horizon agentic workflows, cybersecurity, and scientific research. Google says it uses more tokens than Gemini 4 Argon to reason and verify its output thoroughly.
How much does Gemini 4 Carbon cost per million tokens?
Through the API in Google AI Studio and Gemini for business, the introductory price is $2 per million input tokens and $10 per million output tokens. That matches Gemini 4 Argon. Because Carbon often uses more tokens per task, measure cost per completed task rather than comparing per-token rates alone.
How do I choose between Gemini 4 Argon and Carbon?
Start with Argon for most software engineering and knowledge work. Move a workload to Carbon when it needs multi-step reasoning, extended verification, or open-ended research, and when the fewer-retries benefit is measurable. Run both on a fixed set of your own tasks before deciding.
Why does Carbon use more tokens than Argon?
Carbon spends extra tokens exploring solution paths, reasoning through steps, and running verification loops before answering. That extra work is the reason it can handle harder tasks with fewer errors. The trade-off is that simple tasks get no benefit and can cost more than they need to.
Gemini 4 Carbon is worth testing on the tasks that have been failing or retrying in your current pipeline. Pick one hard workflow this week, run it through Carbon and Argon side by side, and log tokens, retries, and accepted outputs. Then share the results with the engineer who keeps rerunning that same failing job.