5 November 2026Published by Silicon Valley InnovationOwnership & standardsStandards
Technology news from BangkokBANGKOK / TH + EN
SV News
ไทย

AI models

Gemini 4 Carbon tops AI rankings, but you can't use it yet

An AI that finds security holes, tests them and writes the fix on its own, and Google is releasing it step by step, starting with defenders.

Sources

อ่านภาษาไทย

A black phone screen covered in rows of purple star-shaped Gemini app icons
Gemini app icons on a phone. Illustrative; Carbon is not in the app.Photo by Solen Feyissa on Unsplash
Jump to a section
  1. Carbon puts Google back in first place, by a single point
  2. Find the hole, test it, fix it, all in one model
  3. Cyber defenders go first; the public has no date yet
  4. Same price as Argon at first, then double
  5. More right answers, but more made-up ones too
  6. Thais can try its older sibling, Argon, free if they sign up by 2 November
  7. FAQ
  8. Sources

Teams inside Google have let a new AI work for days at a stretch, chasing the root cause of a bug buried in a large codebase, pushing past dead ends until it fixed it.

That AI is Gemini 4 Carbon, which Google DeepMind announced on 28 October 2026, US time, just two weeks after Gemini 4 Argon. Koray Kavukcuoglu, a senior vice president at Google DeepMind, wrote in Google's blog post that it is the company's strongest model, especially for cybersecurity and long-running work.

But if you want to try it today, you can't. Google is giving vetted cyber defenders first access.

Carbon puts Google back in first place, by a single point

On the AI intelligence index run by Artificial Analysis, an independent tester (Intelligence Index v5.0), Carbon scored 59, ahead of Anthropic's Claude Opus 5.5 at 58 and OpenAI's GPT-6 Astra at 57. Argon, the model before it, scored 50, VentureBeat reported.

A gap of one or two points means the three big labs are running almost neck and neck, and each still leads somewhere. Claude Opus 5.5 leads the OSWorld computer-use test at 87.2% against Carbon's 81.5%, and GPT-6 Astra leads software engineering on FrontierSWE 2 at 65.5% against 61.2%.

The test closest to a business owner's world is Andon Labs' Vending-Bench 2, which has an AI run a simulated vending machine business for a year. Carbon ended the year with USD 16,484 (about THB 555,000), ahead of GPT-6 Astra at USD 15,515 and Argon at USD 13,718, VentureBeat reported.

Find the hole, test it, fix it, all in one model

Picture a locksmith who walks through a whole house, finds which doors don't lock properly, tries whether they can be forced open, then fits new locks. Google measures Carbon's security work in the same three steps.

  1. Finding holes: Carbon scored 88.2% on SEC-bench Pro, which measures finding real vulnerabilities in software.
  2. Testing whether a hole is really dangerous: 54.8% on ExploitGym, up from Argon's 42.0%. This step tells a team which holes to close first.
  3. Fixing the hole at its root: 71.0% on CWE-bench.

The real-world example comes from Google's own Chrome team, which says Carbon wrote security fixes more than 10 times better than a large model from another lab, without naming the lab. All the numbers in this section are Google's own and were run without guardrails, so results in real use may differ.

Cyber defenders go first; the public has no date yet

The same skills that help an IT team close holes can also find holes to break in, and Google is opening Carbon up step by step.

This week it goes to a group of cyber defenders, then to more trusted partners through the Fairwind Program, which gives priority to governments and national cyber authorities, critical infrastructure operators and core technology platforms. Applicants go through background checks and may use the model only for defensive work.

Developers, businesses and consumers will get it "as soon as possible", Google says, after it gathers feedback and refines its guardrails, but there is no date. Google is also putting Carbon through the US government's voluntary pre-release model testing.

Same price as Argon at first, then double

Google has already set developer prices, even though the model is not generally available. At launch, Carbon costs USD 2 (about THB 67) per million input tokens and USD 10 (about THB 337) per million output tokens, the same as Argon. A token is the unit AI counts text in, roughly a word or part of a word. After the introductory period, the price rises to USD 4 and USD 20.

For example, a job that feeds in 10 million tokens and gets 1 million tokens back would cost about USD 30 (about THB 1,011) at launch prices and USD 60 (about THB 2,022) afterwards, by VentureBeat's calculation.

More right answers, but more made-up ones too

There is a catch.

On Artificial Analysis's AA-Omniscience test, which checks whether an AI really knows something or makes it up, Carbon answered 58% correctly, up from Argon's 48%. But its hallucination rate, how often it makes up an answer, also rose from 4% to 11%, VentureBeat reported. Google says its internal tests show Carbon makes about half as many unsupported claims.

Work that cannot be wrong, such as figures in financial statements or legal points, still needs a person to check it every time, whichever AI you use.

Thais can try its older sibling, Argon, free if they sign up by 2 November

Carbon isn't here yet, but Thais can already use Argon for free.

Thailand's Ministry of Digital Economy and Society and Google are offering 500,000 free Gemini Enterprise licences to learners on AiPASS, for up to 10 months. Learners need 1,000 points from its courses and must claim by 2 November 2026, or before the licences run out, with use starting on 9 November, The Standard reported. The programme page says the licence includes Gemini 4 Argon.

What you can do now:

  1. Complete 1,000 points of courses on AiPASS before 2 November and claim a Gemini Enterprise licence to try long jobs on Argon.
  2. If your organisation runs critical systems, such as a hospital, a bank or a telecoms network, read the Fairwind Program's terms and fill in the access form on the programme page.
  3. If you plan to use the API, budget at the post-launch prices of USD 4 and USD 20 per million tokens, so the price rise doesn't catch you out.

FAQ

What is Gemini 4 Carbon?

Google DeepMind's newest AI model, announced on 28 October 2026. It focuses on long-running coding work and cyber defence, and ranks first on Artificial Analysis's Intelligence Index v5.0.

Can the public use Gemini 4 Carbon yet?

No. The first release goes to vetted cyber defenders, then widens through the Fairwind Program. Developers, businesses and consumers have no date yet.

How much does Gemini 4 Carbon cost?

At launch, USD 2 per million input tokens and USD 10 per million output tokens. After that, USD 4 and USD 20.

How can Thais use Gemini 4 Argon for free?

Complete 1,000 points of courses on AiPASS, then claim a Gemini Enterprise licence, which includes Argon, by 2 November 2026 or before the 500,000 licences run out. It is free for up to 10 months from 9 November.

Sources

People, companies & topics

GeminiGooglecybersecurityAI modelsAiPASS

More from SV News

Share to Stories

Save this image or choose an app on your phone. Add the article URL with Instagram's Link sticker.

Preparing image...

Card colour
Card photo
Card text (choose up to 3)

Search

Saved

Saved in this browser

Your reading list starts here. Save a story to come back to it later.

Story photography