Gemini 4 Carbon review: pay for it on the hardest step of a job, and keep Argon on the rest
Google's new top model costs the same per token as Argon until November 30, then the price doubles. The bigger difference is in Google's own docs: Carbon thinks longer on every turn, so the same job uses more tokens and takes more time.

Gemini 4 Carbon, generally available since November 10, 2026, is worth paying for on the hardest step of an AI job: root-cause debugging, architecture review, dense document synthesis and security audits. Google prices it like Argon, at $2 per million input and $10 per million output tokens through November 30 and $4 and $20 from December 1, but Carbon spends more thinking tokens per turn. Keep Argon on routine work and give Carbon only the steps where a wrong answer is expensive.
What Google released on November 10
Gemini 4 Carbon is Google's strongest model, and as of November 10 it's open to paying developers and businesses through the Gemini API, Google AI Studio and Gemini for business. Google AI Ultra subscribers also get it in the Gemini app, AI Mode in Search and Google Antigravity. Two weeks earlier it went only to a group of cybersecurity defenders through Google's Fairwind Program.
Google is clear that Carbon doesn't replace Argon, the model it launched in October. In the launch post, Argon "remains the ideal choice for the majority of software engineering and knowledge work," while Carbon is for "tasks that test the limits of a model's intelligence," and it uses more tokens than Argon to "explore thoroughly, reason deeply, and run extended verification loops."
The practical facts for a buyer, as published on November 11:
- Price: $2 per million input tokens and $10 per million output tokens through November 30, 2026. From December 1, that becomes $4 and $20. Cached input drops from $0.10 to $0.20 per million at the same time.
- Thinking is billed. Every thinking token counts as output. Google also bills tool instructions and function declarations when you use grounding, and Search grounding is billed per query.
- No free tier for Carbon.
- Context: a 1-million-token window shared by input, thinking and output, with three thinking levels: low, medium (the default) and high.
Who it is for
Teams that already run AI inside a product or an internal workflow and have one step that keeps going wrong. The bug that only shows up under load. The contract set where clause 14 contradicts an appendix. The security review nobody has time for. Those are the jobs Google names, and they're the right ones.
It's less useful if your AI work is mostly routine: answering order questions, summarizing tickets, tagging products. Argon or a cheaper model already does that well, and Carbon would do it more slowly for more money.
One more group to separate out: people who mainly want to chat with it. Carbon is rolling out in the Gemini app and AI Mode for Google AI Ultra subscribers, which is a consumer subscription. Putting it inside your own product or workflow means the paid Gemini API, with per-token billing and the December price change.
Decision guide
Carbon and Argon, as Google describes them
| Consideration | Gemini 4 Carbon | Gemini 4 Argon |
|---|---|---|
| Google's suggested use | Complex root-cause debugging, architectural design, dense multimodal synthesis, scientific proofs, security auditing | Most advanced software engineering and knowledge work, interactive coding, fast production loops |
| Thinking per turn | More thinking tokens, longer latency | Fewer thinking tokens, faster responses |
| Price per million tokens to Nov 30, 2026 | $2 input, $10 output (thinking billed as output) | Same per-token price, per Google's docs |
| Price from Dec 1, 2026 (Carbon) | $4 input, $20 output, $0.20 cached input | Check Argon's own pricing page before you budget |
| Free tier | Not available | See Google's pricing page |
Same price per token is not the same price per job.
Same price per token, different price per job
This is the part of the launch most people will skim, and it's the part that decides your bill. Google's developer docs say Carbon and Argon share "the same 1M unified context window, built-in tools, and per-token pricing," and that "the main difference is how many thinking tokens each model spends per turn." Carbon spends more, "which increases per-turn token use and latency compared to Argon."
Same price per token is not the same price per job.
Whatever a hard debugging request costs you on Argon today, expect the same request to cost more on Carbon, because the model thinks for longer before it answers, and you pay for that thinking at the output rate. Then the per-token price doubles on December 1. A team that tries Carbon in November and budgets from that month's invoice will be surprised in January.
Google's docs give the right test: run your evaluations at all three thinking levels "then pick the lowest level that meets your quality bar." Do that before you choose a model, not after.
There's also a cap you control. Google's docs say max_output_tokens caps thinking and output combined, and they tell existing apps with a limit under 32K to raise it to at least 32,768. So the cap protects your bill, but set too low it can cut a long answer short. Set it per step, not once for the whole app, and log how often a request hits the limit. If it happens a lot, that step may be a better fit for a person than for a bigger model.
Where it is strong
Long, hard reasoning. Google says Carbon runs build-and-verify loops, checks both code and the rendered interface, and keeps going through dead ends. Its claimed benchmark results include first place on CursorBench 4.0 and Zapier's AutomationBench at less than half the cost per task of the next best external model. These are Google's figures, but they point at the same job: long, multi-step work.
Dense documents. Google says Carbon follows strict output schemas and has fewer unsupported claims than Argon, roughly half as many on Google's internal tests. For legal, finance and operations teams turning piles of PDFs into structured data, that's the property that matters.
Security review. Carbon's cybersecurity results were the headline at its October announcement. For most businesses, the realistic use is a second reviewer on code before release, not running your own vulnerability research.
Easy switch from Argon. Moving an Argon request to Carbon needs no schema change. You change the model name, raise your output limit and timeouts, and test.
How the work moves
Giving Carbon one step, not the whole job
- Sort and gather
Argon, or a smaller model, reads the request and pulls the documents and logs.
- Decide where it's hard
A rule flags the cases that need deep work: a bug nobody can reproduce, a contract with conflicting clauses.
- Carbon on the hard case
Carbon gets the flagged case at the lowest thinking level that passed your tests.
- Check and record
A person or an automated test checks the result; the cost of the run is logged with it.
Where it falls short
Slower answers. Google itself says to keep Argon "where speed matters." A customer-facing chat that waits several minutes is a broken chat. Google's guidance is to allow at least 15 minutes for hard prompts, or to run them in the background.
Budgets get harder to predict. Because thinking length varies by request, the same kind of job can cost very different amounts. You need a spending limit and a log of what each run cost.
Benchmarks are Google's. Every number in the launch comes from Google or its partners. None of them used your tickets, your codebase or your documents.
Migration details. If you're coming from Gemini 3 or earlier, explicit context caching isn't supported, and temperature settings are ignored. Small things, but they break pipelines that assume them.
Plain verdict
Gemini 4 Carbon is a strong model for the one hard step inside a job, and Google's own docs are honest that it isn't the default for everything. Use it for the expensive mistakes: the root cause, the contract conflict, the security pass. Keep Argon, or a smaller model, on the steps around it.
Before November 30, take fifty real cases from the hardest step you have, run them on Argon and on Carbon at low and medium thinking, and record the result and the cost of each run. Price the winner at December's rate. Our Gemini 4 Argon review covers the same December budgeting problem for the cheaper model, and setting an acceptance gate before a model takes the job walks through the scoring.
If you're building an agent and want help deciding which steps deserve the expensive model, that's our AI applications and agents work. Talk to us.
Related work: Spin:Market
We build alongside the people doing the work.
Bring us the problem you’re thinking about.