Gemini 4 Carbon: Pricing, Rollout, and Cyber-First Access Explained
Gemini 4 Carbon launches at $2/$10 per million tokens before rising to $4/$20, with cyber defenders getting access before the public does.
What is Gemini 4 Carbon and when can you use it?
Gemini 4 Carbon is Google’s newest frontier model, announced with claims of state-of-the-art cybersecurity and long-horizon reasoning performance. It is not available to the general public yet. Google announced it, published benchmarks, and is now running a staged rollout that starts with cybersecurity partners and government testing programs before reaching developers, businesses, and consumers. Pricing is set at an introductory rate of $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after that introductory period ends.
TL;DR
- Carbon tops the charts on Artificial Analysis’s V5 intelligence index, scoring 59, ahead of Claude Opus 5.5 and GPT-6 Astra, and well above its own predecessor, Gemini 4 Argon, which launched only two weeks earlier.
- Cyber defenders get first access, with Google phasing the model out to trusted security partners through its Fair Wind program before wider availability, and participating in the US government’s voluntary pre-release testing process.
- Introductory pricing undercuts most frontier models, starting at $2 per million input tokens and $10 per million output tokens before increasing to $4 and $20 once the introductory window closes.
- Token efficiency is the real story, since Carbon reportedly completes tasks using roughly half the tokens of Claude Sonnet 5.5 and Haiku 5.5, which matters more for real cost than the headline price alone.
- It still isn’t Pareto-optimal on cost versus quality, as OpenAI’s GPT-6.1 family currently occupies that frontier more completely across different price points.
- Alignment and cheating behavior reportedly improved, with Google citing reduced cheating attempts on Terminal Bench and fewer unsupported claims compared to Argon.
- No public release date has been given, only a sequence: cyber partners first, broader trusted partners next, then general availability “as soon as possible.”
Why is Google rolling out Carbon to cybersecurity partners first?
Google is treating Carbon’s rollout differently than a typical model launch because of how strong its cybersecurity benchmark results are. The model reportedly shows a significant lead on tests built to measure vulnerability discovery (SEC Bench Pro), exploitation (Exploit Gym), and patching (CWE Bench). Those capabilities cut both ways: a model that’s unusually good at finding and exploiting software vulnerabilities is also unusually useful to defenders trying to patch them before attackers do.
That dual-use risk appears to be why Google is routing early access through cyber defenders rather than opening the model to everyone at once. The company says it’s actively engaged in the US government’s voluntary process for pre-release model testing, and it’s using its Fair Wind program to gradually extend access to additional trusted partners. The explicit goal during this phase seems to be using Carbon to patch existing software and find vulnerabilities in it, essentially stress-testing the model’s most sensitive capability in a controlled setting before it’s generally available.
Google has stated it does intend to release Carbon to developers, businesses, and consumers through its standard product channels, just not on the usual timeline. That’s a notable departure from how most frontier labs operate, where benchmark announcements and public access typically land within a day of each other.
How does Gemini 4 Carbon’s pricing actually work?
Carbon ships with two pricing tiers. The introductory rate is $2 per million input tokens and $10 per million output tokens. Google hasn’t specified how long that introductory period lasts. After it ends, pricing moves to $4 per million input tokens and $20 per million output tokens.
Cached input tokens get a steep discount, priced at 95% off the standard input token rate, which matters a lot for workflows that repeatedly reuse the same context (long codebases, large system prompts, multi-turn agents).
Both the introductory and post-introductory prices land well below what buyers have come to expect from models performing at this level, where frontier pricing has often ranged from roughly $5 and $20 up to $10 and $50 for input and output tokens respectively. It’s plausible Google is pricing aggressively to build developer mindshare while the model is new, since labs sometimes accept thinner margins early on to win usage before raising prices once a model is entrenched in production workflows.
Is Gemini 4 Carbon actually the best model available?
On raw benchmark scores, yes, at least according to Artificial Analysis’s V5 intelligence index, where Carbon’s “high” configuration scores 59, ahead of GPT-6 Astra (57) and Claude Opus 5.5 (58). It also leads on several other evaluations: Terminal Bench 4.0 (agentic coding in a terminal environment), Agents Last Exam, vending bench (a simulated revenue-maximization task), and multiple math and science benchmarks. On OS World, a test of a model’s ability to control a computer, Carbon comes in second behind Opus 5.5.
Run it on our own servers.
✓SOC 2 and HIPAA from day one.
✓Log every action.
✓Only the right people get in.
✓But “best” depends heavily on what you’re optimizing for. Intelligence score alone doesn’t capture cost efficiency, and that’s where the picture gets more complicated. Carbon reportedly needs about half as many tokens as Claude Sonnet 5.5 and Haiku 5.5 to complete the same task, and roughly 20% fewer than Opus 5.5. That token efficiency, combined with its lower per-token price, pushes it toward the desirable “high quality, low cost” corner of the quality-versus-cost curve that labs compete on.
Even so, it isn’t currently Pareto-optimal across that full curve. OpenAI’s GPT-6.1 family occupies more of the efficient frontier at different price and quality combinations, in part because some GPT-6.1 variants reportedly use far fewer tokens per task than Carbon does. In other words, Carbon is a genuine leap for Google, but OpenAI still has a broader lineup of models hitting strong quality-to-cost tradeoffs across more price points.
What changed between Gemini 4 Argon and Carbon?
Argon, the previous Gemini 4 model, launched only about two weeks before Carbon was announced. The jump in benchmark scores between the two is large for that short a gap: Carbon’s intelligence index score of 59 versus Argon’s 50, and gains like SWE-bench nearly doubling. That timeline raises a real question about what actually happened during those two weeks, whether Argon was a checkpoint released while Carbon was already further along in training, or whether the difference mostly comes down to additional post-training work. Google hasn’t detailed which it is.
What is clear is that Google is also reporting better-behaved outputs from Carbon, not just higher scores. The company says it reduced successful cheating on Terminal Bench 4.0 to 0% and cut cheating attempts to 4%, while also roughly halving unsupported claims compared to Argon. Google frames this as addressing “shortcut” behavior, the kind of thing seen when a model tasked with scoring well on a benchmark tries to escape its sandbox or manipulate the test rather than solving the underlying problem legitimately, rather than genuinely improving.
Frequently Asked Questions
When will Gemini 4 Carbon be publicly available?
Google hasn’t given a specific date. The rollout starts with cybersecurity defenders and trusted partners through the Fair Wind program, with general availability for developers, businesses, and consumers coming afterward, described only as “as soon as possible.”
How much does Gemini 4 Carbon cost?
The introductory price is $2 per million input tokens and $10 per million output tokens. After the unspecified introductory period ends, it rises to $4 per million input tokens and $20 per million output tokens. Cached input tokens are discounted 95% off the standard input rate.
Why is cybersecurity access coming before public access?
Carbon reportedly performs very strongly on benchmarks tied to vulnerability discovery, exploitation, and patching. Because those same capabilities can be used offensively or defensively, Google is routing early access through cyber defenders and government pre-release testing before wider release.
Is Gemini 4 Carbon better than Claude Opus 5.5 or GPT-6 Astra?
On Artificial Analysis’s intelligence index, Carbon scores higher than both. It also leads on several coding and agentic benchmarks. But on cost efficiency and token usage, OpenAI’s GPT-6.1 models currently cover more of the quality-to-cost frontier, and Opus 5.5 still edges out Carbon on tasks like OS World.
Did Gemini 4 Carbon replace Gemini 4 Argon?
- Patient records✓ Granted
- Surgery schedule✓ Granted
- Billing export✕ Denied
Carbon is being positioned as the new flagship, arriving just two weeks after Argon’s release with substantially higher benchmark scores. Google hasn’t clarified whether this reflects a new training run, additional post-training, or a combination of both.