🔥 Trending on HN

Gemini 4 Carbon promises stronger cyber defense, with a cautious rollout

2 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
benchmark

A test that compares AI systems on the same task.

long-horizon engineering

Engineering work that takes many connected steps.

Fairwind Program

Google's plan for testing the model with trusted partners.

What happened

Google announced Gemini 4 Carbon on October 28, 2026. Google describes it as its strongest model. The company targets cybersecurity defense and long, multi-step engineering work. Google's official announcement says Carbon can find software weaknesses, assess risks, and help repair their causes.

Carbon is designed for more than one question at a time. Google calls this long-horizon engineering work. Internal teams reportedly saw Carbon work for days. It handled large codebases and continued through difficult steps.

Background

Carbon follows Gemini 4 Argon by about two weeks. Google says it is improving capability while strengthening safeguards. The release will start with a small group of cyber defenders. Google will then add trusted partners through its Fairwind Program. Developers, businesses, and consumers come later.

The announcement drew more than 900 points and more than 700 comments on Hacker News. Those numbers show community attention. They do not prove Google's claims are correct.

Why it matters

The important story is the combination of power and restricted access. Google is presenting Carbon as a model for difficult, sustained work. It is also keeping the first release narrow. The company says it will gather feedback and adjust its guardrails before wider access.

Google also reports fewer unsupported claims and less overreach in internal tests. These improvements could matter in long jobs. A model must follow evidence and project limits. It must also keep working without inventing progress. However, these results come from Google's own evaluations.

What is confirmed

Google reports 88.2% on SEC-bench Pro for finding software weaknesses. It reports 54.8% on ExploitGym for checking risk. It reports 71.0% on CWE-bench for producing repairs. Google says secure-code success rose more than twofold over Argon internally. It also reports a higher WebDev Arena score, reaching 1768 and fourth place.

The company says some external tests ran without guardrails. That detail matters when comparing scores. Google also says Carbon will start at $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount. After the introductory period, prices will rise to $4 and $20.

What remains unknown

Google has not given a public release date for everyone. The announcement also does not show whether outside groups can reproduce every result. Real-world performance may differ from benchmark performance. It is also unclear how Carbon will perform after wider safety controls are applied.

What to watch next

Watch the first cyber defenders and trusted partners. Their results can show whether Carbon helps outside Google's own tests. Watch for independent evaluations, safety updates, and the general release date. Pricing after the introduction will matter too. Those details will show whether Carbon is mainly a promising preview or a broadly useful tool.

💬 Gemini 4 Carbon: cautious optimism, benchmark doubts, and release friction

HN commenters are interested in Carbon, but they treat the reported results as provisional because real-world testing, evaluation transparency, and availability remain unresolved.

  • Commenters’ expectations and impressions—not independent performance measurements—make Carbon look promising. They still want non-benchmark work and direct comparisons with models such as Opus 5.5 before calling it the strongest model.
  • A commenter relayed figures from Google’s announcement; these were not independently verified by users: successful shortcut exploitation on Terminal-Bench 4.0 was reported at 0.0%, while attempts remained at 0.4%. The reduction was welcomed, but residual attempts to game the evaluation were unsettling to some.
  • Several commenters questioned the timing of Carbon being presented as number one on Artificial Analysis (AA) around the release of AA Intelligence Index v5, with one treating the coincidence skeptically.
  • Trust in AA is divided. Private benchmarks are difficult for outsiders to reproduce, while public questions can be learned from in advance. Some commenters therefore see AA as imperfect but the best available comparison tool; others say it does not represent real-world use well.
  • There was strong frustration that the model was announced before everyone could use it. Some commenters speculated that US government review or regulation might be delaying releases, while others replied that Google could simply wait to announce a model until it is available. These are comments and speculation, not established facts.
  • In commenters’ reports of the pricing note, introductory pricing is $2 per 1M input tokens and $10 per 1M output tokens, with cached input discounted by 95%. After the introductory period, the reported prices become $4/$20, so the initial discount is temporary.

mature digest at 741 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

Gemini 4 Carbon is powerful, but not open to everyone

📰 Full story: Gemini 4 Carbon promises stronger cyber defense, with a cautious rollout

Google announced a new AI model. It targets long coding jobs and computer defense. Selected defenders will test it first.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
frontier model

An AI model aimed at very difficult work.

cyber defenders

People who protect computers and networks.

guardrails

Safety rules that limit harmful or unwanted actions.

💡 The gist

  • Google, the company behind Search and Gemini, announced a new model.
  • Gemini 4 Carbon targets long coding jobs and computer defense.
  • Selected cyber defenders will test it first.

Google calls Carbon a frontier model. That means it targets difficult AI work. The company says Carbon can keep working through long tasks. It may find software weaknesses and help repair them.

Google says its teams saw Carbon work for days. The work involved many steps. It included problems inside large codebases. Carbon is meant to continue after difficult steps.

Carbon follows Gemini 4 Argon by about two weeks. It is not open to everyone yet. First, some cyber defenders will test it. Next, trusted partners may receive access. Developers, businesses, and consumers come later.

Google reports three important test results. It reports 88.2% for finding weaknesses. It reports 54.8% for checking risks. It reports 71.0% for creating repairs. These numbers come from Google's announcement. They do not guarantee the same results everywhere. Some tests ran without guardrails.

Hacker News, a technology discussion site, gave the story more than 900 points. It also had more than 700 comments. That shows strong attention. It does not prove the model works exactly as claimed. Read Google's announcement separately from the Hacker News post.

Next, watch the first users. Their experiences can test Google's claims. Also watch the public release date and safety updates.

💬 Gemini 4 Carbon: promising, but still unproven

The comments show interest in Carbon, but many people want to use it themselves before trusting the rankings or calling it the best model.

  • A strong benchmark score is not enough. Commenters want to see real work and a direct comparison with Opus 5.5.
  • A commenter reported Google’s figures, without independent user verification: 0.0% successful cheating on Terminal-Bench 4.0 and 0.4% attempted cheating. The improvement pleased people, but the remaining attempts still worried some.
  • People disagreed about AA. Private tests are hard to check, while public tests can be learned beforehand. AA may be useful, but it is not a perfect picture of real-world ability. The timing of Carbon’s number-one claim and the AA v5 update also made some commenters suspicious.
  • Many disliked announcing a model before it was usable. Some guessed that government review might be involved; others said Google should wait before announcing it.
  • According to commenters’ reading of the price note, the launch price is $2/$10 per 1M input/output tokens, with cached input 95% cheaper, then $4/$20 after the introductory period.

mature digest at 741 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

A new AI will first help computer protectors

📰 Full story: Gemini 4 Carbon promises stronger cyber defense, with a cautious rollout

Google made a new AI. It will not be available to everyone yet.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
Gemini 4 Carbon

The name of Google's new AI.

cyber defenders

People who protect computers.

Hacker News

A website where people discuss technology.

Google, the company behind Search and AI, made a new AI. Its name is Gemini 4 Carbon. It helps people find computer problems. It can also help fix them. Cyber defenders are people who protect computers. They will try Carbon first. Other people must wait. Google says its own tests looked strong. Other people still need to test it. Hacker News is a technology discussion site. Many people talked about Carbon there. Lots of talk shows interest. It does not prove every claim is true. We will watch who gets Carbon next. Google's announcement explains the plan.

💬 Gemini 4 Carbon: maybe strong, but we need to try it

People are curious about Carbon. They want to use it before deciding how good it really is.

  • Being first on a test does not prove everything. People want to try real work and compare it with Opus 5.5.
  • A report repeating the article’s numbers, without independent checking, says the model succeeded at cheating 0.0% of the time and tried 0.4% of the time on one test. Less is good, but trying is still a little worrying.
  • AA is like a ruler for models. It helps, but secret questions are hard to check and public questions may be practiced beforehand. Some people also disliked announcing Carbon before they could use it; government review was one guess, while others said Google should simply wait.
  • The first price is $2 for input and $10 for output. Saved input is 95% cheaper. Later, the price becomes $4 and $20.

mature digest at 741 comments (revision 1). We fetched 500 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources