Google announced Gemini 4 Carbon on October 28, 2026, two weeks after Gemini 4 Argon. Google DeepMind SVP Koray Kavukcuoglu calls Carbon the company’s strongest frontier model, with state-of-the-art cybersecurity and long-horizon coding claims. Artificial Analysis puts it at number one on Intelligence Index v5.0 with a score of 59.
Most readers cannot use it yet. Google is opening Carbon this week to a subset of cyber defenders, then to trusted partners through its Fairwind Program. Developers, businesses and consumers come later, with no date given. Argon remains the Gemini 4 model the public can already reach. We covered the gated Fairwind start in our October 1 Argon post and the wider rollout in our October 14 Argon global launch post.
What Google Claims On Cyber And Long-Horizon Work
On Google’s blog, Carbon leads three external cyber benchmarks that Google says were run without guardrails. SEC-bench Pro scores 88.2% for finding vulnerabilities. ExploitGym scores 54.8% for exploitation. Collinear’s CWE-bench scores 71.0% for patching. Google also says Carbon more than doubles Argon’s success rate at generating secure code on internal benchmarks, a point Channel News Asia and Reuters repeated from the same release.
Google describes Carbon as built for sustained agentic workflows that can run for days across large codebases. It reports a jump of 89 Elo points to number 4 (1768) on WebDev Arena. Safety notes on the same post include 0.0% successful cheating on Terminal-Bench 4.0, with cheat attempts at 0.4%, and a 0.0% attack success rate on Gray Swan’s Indirect Prompt Injection benchmark, down from Argon’s 0.7%.
Number One On Artificial Analysis, With Gaps On Google’s Own Chart
Artificial Analysis scores Gemini 4 Carbon (high reasoning) at 59 on Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Gemini 4 Argon at 50. VentureBeat notes that Carbon posts the highest result in 15 of the 17 rows in Google’s announcement chart. OpenAI’s GPT-6 Astra still leads one software engineering row. On FrontierSWE 2, VentureBeat reports Astra at 65.5% and Carbon at 61.2%. Anthropic’s Claude Opus leads one computer-use row. On the OSWorld-2.1 offline subset, VentureBeat reports Opus at 87.2% and Carbon at 81.5%.
Those 15 of 17 wins and the named Astra and Opus gaps come from VentureBeat’s reading of Google’s chart. Google’s blog itself highlights the cyber scores and the Artificial Analysis lead without listing every rival row.
Artificial Analysis Hallucination Tradeoff
On AA-Omniscience v1.1, Artificial Analysis gives Carbon the top Index score of 54, with accuracy at 58% correct answers versus Argon’s 48%. The same evaluation shows Carbon’s hallucination rate at 11%, up from Argon’s 4%. Artificial Analysis defines that rate as incorrect answers divided by incorrect, partial and unanswered responses, so 11% is not 11% of every answer. Artificial Analysis still ranks Carbon’s hallucination rate second-lowest among models it lists, behind only Argon, while Claude Opus 5.5 (max) and GPT-6 Astra (max) sit at 50% on that component.
Google’s blog reports a different internal measure. It says Carbon cuts unsupported claims by about 2x over Argon in internal testing. That company figure and Artificial Analysis’ Omniscience hallucination rate are separate tests. Do not merge them into one reliability story.
Pricing At Time Of Writing
Google says Carbon launches at the same introductory API price as Argon. At time of writing that is $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off the input price. After the introductory period, Google says $4 per million input tokens and $20 per million output tokens will apply. Google has not stated when Carbon’s introductory period ends. Argon’s higher rates take effect December 1, 2026, per Google’s Argon developer materials, so treat Carbon’s promo window as open-ended until Google names a date.
Artificial Analysis charts Carbon at the standard $4 and $20 rates and notes Google’s 50% launch discount with no confirmed end date. For budgeting, plan on the post-intro $4 and $20 figures unless Google publishes a Carbon end date.
Access Path This Week
Google says it is in the U.S. government’s voluntary pre-release testing process. Starting this week, a subset of cyber defenders gets Carbon. Access then expands to additional trusted partners through Fairwind. Broader developer, business and consumer surfaces come “as soon as possible,” with no calendar date. Artificial Analysis marks the model as having no publicly available API. Channel News Asia’s Reuters-sourced report repeats the cyber-first Fairwind path and the matching intro token prices.
If you need a Gemini 4 model you can call today, stay on Argon. Carbon’s scores are for a gated cohort until Google opens a public endpoint.
JamoraquAI Take
Carbon is the second Fairwind-first Gemini 4 drop in a month. The charts look strong, especially on Artificial Analysis v5.0 and Google’s cyber trio, and the intro token price matches Argon. The useful reading is narrower. You cannot treat Carbon as a drop-in upgrade in your stack this week. Cyber defenders and Fairwind partners are the audience Google named. Everyone else waits.
Keep shipping on Argon while you watch for a Carbon GA date. Budget the post-intro $4 and $20 rates. Weigh Artificial Analysis’ higher Omniscience hallucination rate against Google’s cyber and coding claims before you rewrite any security workflow around a model you cannot reach yet.
FAQ
Can Developers Or Consumers Use Gemini 4 Carbon Today?
At time of writing, Google is opening Carbon to a subset of cyber defenders this week, then to trusted Fairwind partners. No developer or consumer date is published.
Does Carbon Beat Every Rival On Google’s Chart?
VentureBeat reports Carbon leads 15 of 17 rows. GPT-6 Astra leads FrontierSWE 2. Claude Opus leads the OSWorld-2.1 offline computer-use subset.
What Does Artificial Analysis Say About Hallucinations?
On AA-Omniscience v1.1, Artificial Analysis reports Carbon’s hallucination rate at 11%, up from Argon’s 4%, while Carbon’s Omniscience Index leads at 54 and accuracy rises to 58% correct.
What Is The API Price?
At time of writing Google lists $2 input and $10 output per million tokens for the introductory period, then $4 and $20. The intro end date for Carbon is not stated.
Sources
- Google blog, Introducing Gemini 4 Carbon, Oct 28, 2026
- VentureBeat, Google unveils Gemini 4 Carbon, Oct 28, 2026
- Channel News Asia / Reuters, Google launches Gemini 4 Carbon, Oct 28, 2026
- Artificial Analysis, Gemini 4 Carbon is the new number 1 on Intelligence Index v5, Oct 28, 2026
- Artificial Analysis, AA-Omniscience v1.1 evaluation page
- Artificial Analysis model page, Gemini 4 Carbon
