Google's new number one tops the independent index, keeps Argon's introductory price, and is available to a security team you are not on. Here is what was published, what was not, and what the missing pieces cost you.
No, and not this month. Google DeepMind announced Gemini 4 Carbon on October 28, 2026, two weeks after Argon, and it is not in the Gemini app, not in Search AI Mode, not in the public Gemini API, and has no published model ID. The public API catalogue still lists Argon as of October 29. What you can read is this: first place on Artificial Analysis Intelligence Index v5.0 at 59, above Claude Opus 5.5 at 58 and GPT-6 Astra at 57 — Google's own cited figure, measured by a third party. Google reports SEC-bench Pro 88.2% (finding bugs), ExploitGym 54.8% (proving exploitability) and CWE-bench 71.0% (patching), all evaluated without guardrails, all vendor-run. Price: $2 per 1M input and $10 per 1M output, cached input 95% off ($0.10) — identical to Argon on day one — then $4 / $20 after an introductory period with no published end date. Access is staged: a subset of cyber defenders first, then more Fairwind partners, then everyone else "as soon as possible", which is not a date. Google did not publish the context window, a model ID, rate limits, regions or modalities. Our 300-runs-a-day model costs $1,800 a month now and $3,600 a month after the increase — same tokens, double the bill, unknown date. Rating: 3.5 / 5. The model looks excellent; the release is an announcement, a price card and a queue.
Gemini 4 Carbon is Google DeepMind's frontier model announced on October 28, 2026 in a post by Koray Kavukcuoglu, SVP of Google DeepMind and Chief AI Architect, and titled, plainly, "our next frontier model". Google describes it as "our strongest model with state-of-the-art cybersecurity and long horizon capabilities". Two things in that sentence are doing the work: strongest, which the index supports, and cybersecurity, which is the reason the release is locked.
Carbon is not a replacement for Argon and Google does not present it as one. The post says Carbon "builds on Gemini 4 Argon, which was launched two weeks ago", and frames the gap as evidence of pace rather than as a generational handover. That distinction matters commercially: Argon remains the model in the public API, in AI Studio, and in the surfaces people can actually reach, while Carbon exists as a capability Google is showing before it is selling it. Read the family as one line with a gap in it — Argon shipped globally on October 13, Carbon was announced on October 28 and shipped to a cohort, and the model the two of them will eventually be compared against does not exist yet.
The capability story is narrower than "frontier model" suggests, and more interesting. Google's cybersecurity claims are structured around the defensive lifecycle rather than a single score: find a real vulnerability (SEC-bench Pro), validate that it can actually be exploited so a team knows what to fix first (ExploitGym), and patch the root cause in working code (CWE-bench). That is the shape of a product for security operations, not a red-team toy, and it explains both the access gate and the government testing sentence in the same paragraph. The long-horizon half of the pitch is the ordinary one: harder engineering problems, sustained agentic workflows, tasks that run for many steps without losing the thread.
What Carbon is not, on the evidence published, is a consumer upgrade. Nothing in the announcement promises a better Gemini app, a faster Search answer, or a cheaper bill for anyone. If you use Gemini to draft email or summarise a PDF, this release does not change your week — which is the subject of its own section below, because a lot of coverage has treated a security model as a general-purpose one.
Stripping the adjectives leaves five categories of claim and a large hole where a specification sheet would be. Here is the whole of it, in the order the post gives it.
Google states that Carbon "ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59". This is the single most credible number in the release, because Artificial Analysis runs its own harness and did not take a Google-provided score. It is also a number that moved the whole field: the v5.0 update re-scored the models this site has been quoting all month, which is why the comparison to earlier reviews needs care. On v5.0, Claude Opus 5.5 sits at 58 and GPT-6 Astra at 57; our own reviews of those models recorded 58 and 53 respectively on v4.3.2, and the Astra figure in particular is not directly comparable across versions.
Artificial Analysis's own write-up is worth reading next to Google's, because it is more specific about where the lead comes from and where it stops. It puts Carbon first in four of the nine index evaluations — HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA — and says plainly that Carbon "makes substantial gains elsewhere but is not leading in coding and knowledge work". On the Artificial Analysis Coding Agent Index it scores 59 with the Antigravity CLI, three points behind Claude Code with Claude Opus 5.5 at max effort at 62, and it is second rather than first on Terminal-Bench 4.0 at 64% and on FrontierCode v1.1 at 51%. On agentic knowledge work it improves a lot and still does not lead: AA-Briefcase 1,609 Elo, a 126-point gain over Argon's 1,483, behind Kimi K3.1 Pro's 1,678 and Muse Wonder's 1,659.
That is a first place in a real competition, not a runaway. The fair reading is that Carbon is the best general model on this index by a single point, that its advantage is concentrated in reasoning, knowledge and science tasks, and that the coding-agent crown is still Anthropic's.
| Benchmark | What it measures | Carbon | Status |
|---|---|---|---|
| SEC-bench Pro | Autonomously finding real vulnerabilities in real-world software | 88.2% | State of the art, per Google |
| ExploitGym | Validating risk by demonstrating exploitability | 54.8% | New state of the art, per Google |
| CWE-bench | Patching at the root cause, turning findings into working fixes | 71.0% | State of the art, per Google |
The parenthetical Google attaches to all three is the most important word in the post: "evaluated without guardrails". These are the model's raw capability figures with its safety systems switched off, which is standard practice for capability measurement and also the reason they say nothing about what a delivered product will refuse to do. A model can score 54.8% on exploitation in an evaluation harness and still decline the same request in production; that gap is a policy decision, not a capability one, and Google is telling you which number it published.
Carbon launches at "the same introductory price as Argon": $2 per million input tokens, $10 per million output tokens, cached input at 95% off. The footnote is the part that matters: after the introductory period the price becomes $4 and $20. Google does not say when that period ends. Argon's own introductory window was published with a date — it expires November 30, 2026 — so the absence of one here is a deliberate omission, not an oversight.
"Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program." Then: "We're continuing to gather feedback and iterate on guardrails before making Carbon available to developers, businesses, and consumers via our product surfaces as soon as possible." And: "We're taking a deliberate approach to rollout and are actively engaged in the U.S. government's voluntary process for pre-release model testing." Three sentences, three audiences, one date — "this week" — and it applies only to the first audience.
This is a launch post of roughly a thousand words that finds room for eight paragraphs on alignment and never states the context window. The omissions are systematic enough to list:
| Item | In the announcement? | Why it matters |
|---|---|---|
| Artificial Analysis v5.0 rank and score | Yes — #1 at 59 | The one third-party number in the release |
| Three cyber benchmarks | Yes — 88.2 / 54.8 / 71.0 | The capability case, unguardrailed, vendor-run |
| Introductory price | Yes — $2 / $10, cache 95% off | Same as Argon on day one |
| Post-introductory price | Yes — $4 / $20, in a footnote | Doubles, on an unstated date |
| Context window | No | You cannot size prompts, chunking or cost without it |
| Model ID | No | You cannot pin a version or write a client |
| Rate limits and quotas | No | You cannot capacity-plan a pipeline |
| Regions and availability | No | You cannot answer a data-residency question |
| Modalities | No | You cannot tell whether it takes images, audio or video |
| Technical report, parameter count, architecture | No | No way to check any claim independently |
| Date for developer or consumer access | No — "as soon as possible" | No way to plan a migration |
Six of the eleven rows are empty, and five of those six are things a developer needs before writing a line of code. That is the case for the 3.5, stated as a table.
There are three audiences in the announcement and only one of them gets something this month. The distinction is not cosmetic: it decides whether Carbon is a tool you can build on or a benchmark you can read.
| If you are… | What the announcement gives you | What you can do today |
|---|---|---|
| A cyber defender in the first cohort | Access "starting this week", through the Fairwind Program, with guardrails still being iterated | Use the model inside the program's terms. No public endpoint, no published quota, and no permission to benchmark it against a public harness and publish the result |
| A trusted partner added later | Gradual phasing, no criteria published, no queue position, no date | Wait. Ask your Google account team. Nothing in the post tells you how a company becomes eligible |
| A developer, business or consumer | "As soon as possible" via product surfaces, after more feedback and guardrail work | Nothing. Not in the Gemini app, not in AI Mode, not in the Gemini API — the catalogue still lists Argon. Use Argon or go elsewhere on this page |
The Fairwind Program is the gate, and it is not new. It is Google DeepMind's defenders-first access programme, launched September 2, 2026 and expanded on October 13 to hundreds of additional partners, with more than 650 participating organisations worldwide at that point. Carbon is the second model to move through it, and the pattern is now clear enough to plan against: Google releases a frontier model to defenders first, watches what they do with it, adjusts guardrails, and then decides about everyone else. Whether that is responsible or self-interested depends on your view, and probably on both.
The uncomfortable part for a normal developer is not the gate, it is the absence of a queue. A gated release with published criteria — apply here, expect a decision in six weeks, these are the requirements — is a delay you can plan around. Carbon's release offers a category ("trusted partners"), a mechanism ("Fairwind"), and an adverb ("gradually"). If your roadmap depends on Carbon, you have been given a direction of travel and no timetable, and the honest planning assumption is that you should not build on it this quarter.
All figures are from Google's October 28 announcement, checked against Artificial Analysis's model page on October 29, 2026, and all are US dollars before tax. Two details make this price card harder to use than it looks: there is no date for the increase, and there is no endpoint to buy the introductory rate on yet. A discount you cannot spend is a marketing number until the model opens.
API — introductory period (no published end date)
$2.00 per 1M input · $10.00 per 1M output · $0.10 per 1M cached input (95% off input)
After the introductory period expires: $4.00 per 1M input · $20.00 per 1M output · cache discount stated as 95% off, so about $0.20 per 1M cached
Exactly double on both legs. Google's footnote gives no date for the change; Artificial Analysis describes the arrangement as a 50% launch discount "with no confirmed end date". For comparison, Argon's introductory window was dated — it ends November 30, 2026 — which makes Carbon's open-ended window the exception in this family.
Security work is not chat, so model it as agent runs. Take a team running 300 runs a day, each sending 60,000 input tokens — a repository slice, scanner output, CVE context and the task instructions — and returning 8,000 output tokens of finding, patch and reasoning.
The arithmetic, written out (no caching)
At the introductory $2 / $10:
Input: 300 × 60,000 = 18,000,000 tokens a day × $2.00 ÷ 1M = $36.00
Output: 300 × 8,000 = 2,400,000 tokens a day × $10.00 ÷ 1M = $24.00
Total $60.00 a day → about $1,800 a month. Per run: $0.20.
At the post-introductory $4 / $20, same tokens:
Input 18M × $4.00 ÷ 1M = $72.00; output 2.4M × $20.00 ÷ 1M = $48.00
Total $120.00 a day → about $3,600 a month. Per run: $0.40.
The difference: $1,800 a month, or $21,600 a year, for identical work, triggered by a date that has not been announced. Budgeting on the headline rate is how a security tooling line item doubles without anyone approving it.
Caching is the lever that changes the shape of that bill, and Carbon's 95% discount is unusually generous. If 80% of the input is a pinned repository and rulebook served from cache: 14.4M cached × $0.10 = $1.44, plus 3.6M fresh × $2.00 = $7.20, plus $24.00 of output, gives $32.64 a day, about $979 a month at the introductory rate — a 46% saving. After the increase the same configuration is $65.28 a day, about $1,958 a month. Caching does not protect you from the increase; it only changes the base the doubling applies to.
| The same 300 runs a day, 60K in / 8K out | Input / 1M | Output / 1M | Per run | Per month |
|---|---|---|---|---|
| Gemini 4 Carbon, introductory | $2.00 | $10.00 | $0.20 | ~$1,800 |
| Gemini 4 Carbon, introductory + 80% cache | $2.00 / $0.10 cached | $10.00 | $0.11 | ~$979 |
| Gemini 4 Carbon, after the increase | $4.00 | $20.00 | $0.40 | ~$3,600 |
| Gemini 4 Carbon, after the increase + 80% cache | $4.00 / $0.20 cached | $20.00 | $0.22 | ~$1,958 |
| Gemini 4 Argon ($2 / $10 until Nov 30) | $2.00 | $10.00 | $0.20 | ~$1,800 |
| GPT-6.1 Terra ($1 / $5) | $1.00 | $5.00 | $0.10 | ~$900 |
| GPT-6 Astra ($10 / $50) | $10.00 | $50.00 | $1.00 | ~$9,000 |
| Kimi K3.1 Pro ($2 / $10) | $2.00 | $10.00 | $0.20 | ~$1,800 |
Our arithmetic at published list prices with the token assumptions above. Your prompts will differ — measure them. Competitor rates approximate, checked October 29, 2026.
Artificial Analysis prices every model at standard rates rather than launch discounts, so its Carbon figure assumes $4 and $20 and reports $7.76 per Intelligence Index task. That sits roughly 23% below Claude Opus 5.5 at $10.04 and about 1.2× GPT-6 Astra at $6.25. Two things follow. First, Carbon is not cheap per completed task even on the index — it earns its place by finishing harder tasks, not by being economical. Second, the same evaluator notes Carbon generated 200 million output tokens to complete the index against a median of 64 million, so it is extremely verbose. At $10 or $20 per million output tokens, verbosity is the hidden line item: a model that talks three times as much costs three times as much on the output leg, whatever the input rate says.
No — and in this case the question is close to moot, because there is nothing to buy either.
Practical read: there is no evaluation budget to set and no trial to start. Wait for the model ID, then price at $4 / $20 as your planning case and treat $2 / $10 as a bonus rather than a baseline.
The awkward thing about recommending alternatives to Carbon is that most of them are worse models. The recommendation is not "buy something better" — it is "buy something you can call", which is a different and more useful question on October 29.
| Model | Input /1M | Output /1M | Cached /1M | AA v5.0 | Who can use it today (Oct 29, 2026) |
|---|---|---|---|---|---|
| Gemini 4 Carbon | $2.00 → $4.00 | $10.00 → $20.00 | $0.10 → $0.20 | 59 · #1 | A subset of cyber defenders via Fairwind; more trusted partners later; no developer or consumer access, no model ID |
| Gemini 4 Argon | $2.00 → $4.00 on Dec 1 | $10.00 → $20.00 on Dec 1 | $0.10 → $0.20 | 53 on v4.3.2 | Gemini API, AI Studio, Antigravity (Ultra), enterprise — generally available |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | 58 | Anthropic API, Claude apps, major clouds |
| GPT-6 Astra | $10.00 | $50.00 | $1.00 | 57 | OpenAI API, ChatGPT Work, Codex, Pro tier |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 | 48 on v4.3.2 | OpenAI API (Build+), ChatGPT Work, Codex, GitHub Copilot |
| Kimi K3.1 Pro | $2.00 | $10.00 | $0.20 | 55 on v4.3.2 | Kimi API, Kimi.ai, Kimi Work, Kimi Code, OpenRouter |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | 38 on v4.3.2 | OpenAI API, ChatGPT desktop app on Free and Go |
Scores in the v5.0 column are from the updated index; figures marked v4.3.2 are what our earlier reviews recorded on the previous version and are not directly comparable — the v5.0 update re-scored the whole field, which is why Astra appears here at 57 and at 53 elsewhere on this site. A model with no published v5.0 figure is not a model that scored zero. Prices in US dollars, checked October 29, 2026; Carbon and Argon both carry an introductory rate that doubles, and only Argon's has a published end date.
Most model launches are consumer products with an enterprise tier attached. This one is the reverse, and it is worth being clear about why, because the reasoning explains the access gate, the price, and the strange feeling of reading a launch post with nothing in it for you.
A model that reaches 88.2% on finding real vulnerabilities and 54.8% on demonstrating exploitation is a tool with two uses and one instruction manual. The same capability that lets a bank's security team triage its backlog lets someone else triage a target's. Google's answer is sequencing: give it to defenders first, watch how it is used, harden guardrails, and open it later. That is a defensible public-interest position and it is also a competitive one — a defenders-first programme builds relationships with exactly the security vendors and enterprises that buy frontier inference at volume, and Google says as much when it notes Wiz and its Chrome team as partners.
The U.S. government's voluntary pre-release testing line in the same paragraph should be read the same way: it is a real constraint on the release schedule and a real signal about how frontier cyber models will be governed from here. Any model family with this capability profile is going to arrive through a government conversation for the next several years.
If the vendor numbers hold, the change is not "AI writes your patches". It is that the triage step gets cheap. The step that costs a security team the most hours is not discovering that a dependency has a CVE — that is automated already — it is deciding which of four hundred findings are real, exploitable, and reachable in your deployment, and then writing the fix. Carbon's three benchmarks map exactly onto those three questions, in order, which is why the claims are more credible than a single aggregate score: they describe a workflow somebody actually has.
The Chrome patch claim — more than ten times better than another lab's frontier model at generating high-quality patches — is the number to watch, because patching is where a wrong answer is expensive and a right one is worth a great deal. If it survives contact with real codebases in the Fairwind cohort, it is the most commercially significant result in the release, and the least verifiable.
Nothing about your writing, your translations, your spreadsheet formulas or your customer support bot. Carbon is not going to make the Gemini app better next week, is not in Search, and is not on a price list you can reach. The models that answer those needs are covered elsewhere on this site, and most of them are cheaper than Carbon's introductory rate, let alone its post-increase rate. The one indirect effect worth noting is competitive: a first place on the index at 59 puts pressure on Anthropic and OpenAI, and pressure on the frontier usually shows up as price cuts in the middle of the market within a quarter or two — which is where ordinary users actually benefit.
This section exists because the release has more unknowns than usual, and because a review that papers over them is not worth reading. Each item below was checked on October 29, 2026 and found unpublished.
| Item | Status | What we advise |
|---|---|---|
| Context window | Not published by Google. Artificial Analysis lists 1M tokens, which is a third-party catalogue entry, not a Google figure | Treat as unknown. Do not size prompts, chunking or cost models around 1M tokens |
| Model ID | Not published. The public Gemini API catalogue lists Gemini 4 Argon, not Carbon | Do not write client code against a guessed string |
| Rate limits and quotas | Not published | No capacity plan is possible; assume tiered limits like other Gemini models once published |
| Regional availability | Not published | Cannot answer a data-residency question for a regulated employer |
| Modalities | Not published. Artificial Analysis lists text, image, speech and video input with text output — again third-party | Assume text-only for planning until Google documents otherwise |
| End of the introductory period | Not published — Google's footnote gives the new price but no date; Artificial Analysis describes the discount as having "no confirmed end date" | Budget at $4 / $20 and treat $2 / $10 as temporary |
| Developer and consumer access date | Not published — "as soon as possible" | Do not put Carbon on a roadmap this quarter |
| The cybersecurity benchmarks | Vendor-reported, not reproducible — the model is not publicly callable, so no third party can rerun SEC-bench Pro, ExploitGym or CWE-bench | Treat 88.2 / 54.8 / 71.0 as claims of capability, not as measured facts about your codebase |
| Technical report, parameter count, architecture | Not published with this release | No basis for a self-hosting or fine-tuning plan; there are no weights |
One further number deserves a note. Carbon's exact v5.0 score is confirmed at 59 by Google and by Artificial Analysis, and Claude Opus 5.5 at 58 is consistent across both. GPT-6 Astra's v5.0 figure appears as 57 in our sources but is not something this site has seen published directly by Artificial Analysis, so treat it as directional rather than exact. Where an outlet and a primary source disagree, this review quotes the primary source.
No. Google announced Gemini 4 Carbon on October 28, 2026 and said it would make the model available starting that week to a subset of cyber defenders, then gradually to additional trusted partners through the Fairwind Program. Developers, businesses and consumers are described in the announcement only as 'as soon as possible', after more feedback and guardrail work, with no date attached. As of October 29, 2026 the public Gemini API model catalogue still lists Gemini 4 Argon rather than Carbon, so there is no model ID to pin, no quota to plan against and no endpoint to test. If your plan for this week involved Carbon, replace it with Argon or with another model from our alternatives list.
In the announcement Google set an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95 percent off, which works out to $0.10 per million cached tokens. A footnote says that after the introductory period expires the price becomes $4 per million input and $20 per million output, exactly double, and Google has published no end date for the introductory period. On a modelled workload of 300 agent runs a day at 60,000 input and 8,000 output tokens each, the bill is $60 a day and about $1,800 a month at the introductory rate, and $120 a day and about $3,600 a month after the increase, for the same tokens. Artificial Analysis prices the model at the standard $4 and $20 rates and reports $7.76 per Intelligence Index task, against $10.04 for Claude Opus 5.5 and $6.25 for GPT-6 Astra.
No. Google has announced no free surface for Gemini 4 Carbon: it is not in the free Gemini app tier, not in AI Mode in Search, and not in the public Gemini API, and the model is closed with no weights published. The introductory rate is a discount for paying API customers, not a free tier, and there is no free quota anywhere in the announcement. The freely usable Gemini models today are the ones already on the public API and in the consumer app, covered in our Gemini review, and Gemini 4 Argon is the current frontier model you can actually call.
Google has not published it. The launch post gives no context window, no maximum output length, no rate limits and no modality list, and no technical report accompanied the release. Artificial Analysis, a third party, lists a 1M token context with text, image, speech and video input and text output, but that is an independent catalogue entry rather than a Google figure and should not be treated as confirmed. Until Google documents the window on its own developer pages, treat the number as unknown: do not size prompts, chunking rules or cost models around 1M tokens for this model, and do not assume any particular modality support.
On Artificial Analysis Intelligence Index v5.0, Gemini 4 Carbon scores 59 and Claude Opus 5.5 scores 58, so Carbon is one point ahead and ranks first, which is the figure Google itself cited. The picture is not uniform. On the Artificial Analysis Coding Agent Index, Carbon scores 59 with the Antigravity CLI, three points behind Claude Code running Claude Opus 5.5 at max effort at 62, and Carbon is second rather than first on Terminal-Bench 4.0 at 64 percent and on FrontierCode v1.1 at 51 percent. Carbon's cybersecurity numbers are vendor-reported and cannot be independently reproduced today, because the model is not on the public API, so nobody outside Google and the first Fairwind cohort can rerun SEC-bench Pro on it.
The most direct substitute is Gemini 4 Argon, the same family, the same introductory $2 and $10 rate and the model that is actually in the public API today, covered in our Gemini 4 Argon review. For cheaper agentic work, GPT-6.1 Terra at $1 and $5 does computer use at a tenth of GPT-6 Astra's token rate. For the most capable model you can legitimately rent, GPT-6 Astra at $10 and $50 and Kimi K3.1 Pro at $2 and $10 with 55 on the earlier index are the usual picks. For hard reasoning with a published model card, Claude Opus 5.5 at $4 and $20 is one point behind Carbon on index v5.0 and ahead of it on coding agents.
Combine multiple PDF files into one — free, no upload, files stay on your device.
📒 Bookkeeping ReviewsHonest accounting software reviews: Wave, QuickBooks, Xero & free options.
🎮 Gaming Gear ReviewsEsports hardware decoded: gaming mice, keyboards, headsets & budget picks.
📷 ID Photo StudioPassport & visa photos with AI background changer, compressor & converter.
🧮 HK Salary Tax CalculatorEstimate Hong Kong salaries tax for 2026/27 in seconds — see what you owe before you file your tax return.
🎁 HK Benefits Quick CheckGovernment & benefits eligibility checker — OAA, CSSA, housing & more.
📅 HK Calendar 香港月曆Public holidays, lunar dates & solar terms for 2020-2036 — bilingual, free, no sign-up.
★ Our own product 🏠 Home Expense BookkeepingFamily bookkeeping: income, spending, net worth and monthly bills. Data stays on your device. HK$98 one-time.