One word from a rival, strong benchmark claims from Google, and no public access yet. The release plan matters as much as the scores.
Gemini 4 Carbon drew a one-word verdict from Elon Musk on October 28, 2026. Musk replied “Impressive,” to a post by Google chief executive Sundar Pichai announcing the model, and Pichai answered with a folded-hands emoji. The exchange was brief.
Google describes Gemini 4 Carbon as its strongest yet, built for cybersecurity defence and long-running engineering work. The company says it can find software vulnerabilities, test whether they can be exploited, and help develop patches. That is a specialist job, not a promise that every user will soon meet the model in a chatbot.
The catch is access. Google is giving the model first to a small group of defenders, then to trusted partners, while it keeps testing safeguards. There is no public release date. Judge it on Google’s evidence, and on what independent testers find in real work.
The Short Version
- Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Argon, in a DeepMind blog post.
- Google says the model ranks first on the Artificial Analysis Intelligence Index v5.0, with a score of 59.
- Google reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, evaluated without guardrails.
- First access: a subset of defenders, then Fairwind Program partners. No wider release date announced.
- Introductory API pricing is $2 input and $10 output per million tokens (USD), rising to $4 and $20 later. No euro (EUR) price has been published.
Gemini 4 Carbon: What Google Announced
The announcement came in a Google DeepMind blog post introducing Gemini 4 Carbon by Koray Kavukcuoglu, Google DeepMind senior vice president and chief AI architect at Google. The post calls the model Google’s strongest, with state-of-the-art cybersecurity and long-horizon capabilities. That wording is a company claim, not an independent finding.
Long horizon is the second half: sustained agent work on multi-step tasks, rather than a single answer. Internal teams, Google says, have seen it run for days, persisting through dead ends while tracking down a bug. The announcement publishes no parameter count, context window, or architecture.
Gemini 4 Carbon Benchmarks, With the Limits Attached
The three headline numbers for Gemini 4 Carbon come from external cybersecurity evaluations run without guardrails, Google says. The model scored 88.2% on SEC-bench Pro (finding vulnerabilities), 54.8% on ExploitGym (proving exploitability), and 71.0% on CWE-bench (patching). “Without guardrails” describes the raw model before safety layers are applied, which is why Google limits who may use it.
Two internal claims sit alongside the external scores. Google says the model achieves more than twice Argon’s success rate at generating secure code. Its Chrome team, Google says, found it performed more than 10 times better than another lab’s frontier model at generating high-quality patches. The other lab is not named and the method is not published, so the figure cannot be checked. Read it as a Google report, not a verified head-to-head result.
On the Artificial Analysis Intelligence Index v5.0, Google says Gemini 4 Carbon ranks first at 59, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57, and Argon at 50. On the same evaluator’s Coding Agent Index, it sits second. A benchmark lead is not proof of real-world performance. PetaTech24 did not test the model and did not verify the scores independently.
Why Gemini 4 Carbon Starts With Defenders Only
Google is rolling Gemini 4 Carbon out in stages. A subset of cyber defenders gets access first, starting that week. Access then expands to trusted partners through the Fairwind Program, Google’s defender early-access route. Wider availability follows after feedback and guardrail work. No date is attached.
Google says it is taking part in the United States government’s voluntary pre-release testing process. The reason for caution is dual use. PetaTech24 analysis: a tool that finds a flaw, proves it exploitable, and writes a fix helps a defender close a hole first. In the wrong hands, it could help an attacker find that hole first.
Google reports Gemini 4 Carbon safety improvements too. In live coding with its own teams, high-severity misalignment alerts fell by almost 80% compared with Argon, and unsupported claims were roughly halved, according to Google. It also reports a 0.0% attack success rate on Gray Swan’s prompt-injection benchmark. All three figures are Google-reported, and no independent safety audit has been published.
The Bigger Picture
The pace is part of the story. Argon arrived two weeks before Gemini 4 Carbon, and Pichai framed it as proof of a periodic release cadence. Each launch is now judged against the previous one. That raises a fair question: can evaluation and safeguards keep up?
Musk’s reaction to Gemini 4 Carbon should be read with the same care. NDTV reported the exchange, noting Musk used the same reply to a Pichai post in July, about quarterly results. A single word is praise, not a review. Musk runs rival AI company xAI, so the praise signals attention. It is not evidence about benchmark quality or safety.
Confirmed, Reported, and Still Unknown
The vendor holds most of the evidence, so the categories must stay apart. Confirmed: the announcement, rollout order, and prices. Reported: the scores and safety improvements. Unknown: whether the claims hold.
Confirmed
- Google announced the model on October 28, 2026, via its blog and Pichai’s post.
- Access starts with a subset of cyber defenders, then expands through the Fairwind Program.
- Introductory pricing: $2 input and $10 output per million tokens (USD), cached input 95% off, rising to $4 and $20 later.
Reported
- Scores of 88.2%, 54.8%, and 71.0% on the three benchmarks, evaluated without guardrails, as reported by Google.
- First place at 59 on the Intelligence Index v5.0, second on the Coding Agent Index, as reported in coverage.
- More than twice Argon’s secure-code success rate, and an almost 80% fall in misalignment alerts, from Google internal testing.
Still unknown
- When wider access arrives, and when introductory pricing ends.
- Parameter count, context window, and architecture, none published in the announcement checked.
- Whether independent testers reproduce the results, and what defenders report from real patching work.
What to Watch Next
Watch first for defender feedback on Gemini 4 Carbon, even in summary form. If teams report real flaws found and patches that survive review, the benchmark story gains weight. If reports stay vague, hold the claims at arm’s length. Watch next for the independent Artificial Analysis breakdown and any reproduction of the scores outside Google.
Watch finally for two missing dates: wider release, and the end of introductory pricing. The first decides when most developers can try the model. The second decides the cost at scale, since standard rates double the introductory figures.
Quick Answers on Gemini 4 Carbon
When can I try Gemini 4 Carbon?
Probably not, unless you are in the first defender group. Access goes first to a subset of defenders, then to Fairwind Program partners. No date has been announced for developers, businesses, or consumers.
What is the model actually for?
Security and long engineering work. Google says Gemini 4 Carbon finds vulnerabilities, tests exploitability, and helps develop patches over days-long tasks. It is a specialist tool first, not a consumer chatbot launch.
How much does it cost?
Introductory pricing is $2 per million input tokens and $10 per million output tokens (USD), cached input at 95% off. Standard pricing is $4 and $20. No euro price has been published, and the introductory end date has not been stated.
Did Elon Musk endorse the model?
He replied with a single word, “Impressive,” to Sundar Pichai’s announcement post. That is praise from a rival AI chief, not a review or an endorsement of any specific claim. Pichai replied with a folded-hands emoji.
The Takeaway
Google has announced Gemini 4 Carbon, published striking benchmark figures, and kept it from the public while defenders test it. That combination is the story: ambition paired with caution, where the same capability can defend or attack. Musk’s one word adds colour, not evidence.
Security teams should watch the defender rollout and independent evaluations, and wait for a public date before planning around the model. Everyone else should treat the scores as Google’s claims until independent testers reproduce them. If the defender evidence holds, the model will matter. Until then, proof is still to come.
Sources and Attribution
- Google DeepMind blog, Koray Kavukcuoglu, “Introducing Gemini 4 Carbon,” October 28, 2026. Primary source for the announcement, benchmark figures, rollout order, pricing, and safety claims.
- Sundar Pichai announcement post on X, October 28, 2026, as quoted in reporting by Digit, TestingCatalog, and others. Source for the release cadence framing and the Intelligence Index score.
- NDTV, “Elon Musk Calls Google’s New Gemini 4 Carbon AI Model ‘Impressive’,” October 2026. Source for Musk’s reply, Pichai’s emoji response, and the earlier July exchange.
- Reuters, via KTWB, “Google launches Gemini 4 Carbon AI model with cybersecurity focus,” October 28, 2026. Corroboration for the launch, the Argon timing, and the Fairwind rollout.
- Artificial Analysis Intelligence Index v5.0 and Coding Agent Index, as reported by Artificial Analysis and in coverage by Dealroom and Wccftech. Source for rival scores and placings.
- The Register, “Google’s Gemini 4 Carbon looks like a diamond in the rough,” October 29, 2026. Corroboration for standard pricing of $4 and $20 after the introductory period.
PetaTech24 tested nothing for this article. All benchmark, safety, and capability claims are attributed to Google or to the reporting named above. Pricing is stated in US dollars only, because no euro price has been published. No parameter count, context window, or general availability date is stated, because none was published in the announcement checked.
