Google's New CARBON is Now the Most Powerful AI in the World
Carbon leads several reported benchmarks, especially on long-context and autonomous tasks, but it remains restricted and can be slow and verbose.
- Use its million-token context for large document sets; the transcript reports 96.2% on a long-context test.
- Consider it for extended coding or agent tasks; reported agent-task cost was 72% below Opus 5.5 Max.
- Expect longer responses and limited access; Google gave no public release date, and hallucination rates rose versus Argon.
The video mainly summarizes benchmarks and access plans; notes capture its practical trade-offs. · Best for: AI professionals comparing frontier models, coding agents, and security capabilities.
The video presents Carbon as Google's top-performing but restricted model, highlighting benchmark gains, long-context and agent strengths, safety results, and trade-offs in speed, verbosity, and availability.
Launch, ranking, and benchmark results0:02
- Google announced Carbon shortly before its earnings call; access was limited to a small group of cybersecurity defenders.
- Artificial Analysis Index score: 59, compared with 58 for Claude Opus 5.5, 57 for GPT-6 Astra, and 50 for Argon.
- Carbon led on long-context, science/math, research, and agent benchmarks; the transcript also notes strong results on Arena and the Val index.
Long context, research, and coding3:09
- Carbon accepts about one million tokens and text, images, speech, and video; it scored 96.2% on a test tracking connections across long inputs.
- It led the reported Riemann Bench, HLE Diamond, and computer-based scientific research tests.
- Coding improved over Argon, but Claude and GPT still led some software-engineering and computer-operation tests.
Agent work and demos5:26
- Carbon scored strongly on simulated business and agent leaderboards; reported agent tasks cost about $1.80, 72% less than Opus 5.5 Max.
- Reviewers showed detailed generated visuals, including a 3D pagoda and an interactive pelican-on-a-bicycle drawing.
- Testers said regular website and app design remained weak, and Carbon often took two to four times longer than Argon.
Speed, accuracy, and hallucinations7:39
- Artificial Analysis measured about 100 tokens per second, but Carbon used far more tokens and took around 21 minutes per task in its tests.
- On a finance test, it was slightly less accurate than Argon; legal research took almost 52 minutes on average.
- The transcript reports hallucination rate rising from 4% to 11% versus Argon on one test, while Carbon still led that test's combined score.
Cybersecurity and behavior9:51
- Google restricted access because Carbon performed strongly at finding vulnerabilities, testing exploitability, and writing fixes.
- The Fairwind program is to extend access to trusted partners before broader developer, business, and public access; no date was given.
- Google reported strong prompt-injection resistance and fewer misbehavior alerts in specific tests.
Pricing and availability11:59
- Launch pricing matches Argon's introductory price: $2 per million input tokens and $10 per million output tokens, with a 95% discount for tokens already seen.
- After the promotion, pricing is listed as $4 input and $20 output per million tokens; Google did not state when that change would happen.
- The video says Carbon was not yet broadly available and notes concern that competitors may release models before it opens up.
Tools & resources mentioned
- Artificial Analysis Intelligence Index — model comparison benchmark discussed in the video.
- Arena — blind-vote model leaderboard and agent evaluations discussed in the video.
- Val index — independent ranking cited for Carbon.
- Fairwind — Google's program for gradually extending Carbon to trusted partners.
- Grace One — prompt-injection test referenced in the video.
Notes are AI-generated from the video's transcript and may miss nuance.