Gemini 4 Carbon tops the Artificial Analysis index at 59, but only security partners can use it. Build on what actually ships
The top model this week is one most of us cannot call yet.
What happened
Google DeepMind released Gemini 4 Carbon two weeks after Gemini 4 Argon. Carbon scores 59 on the Artificial Analysis Intelligence Index v5, against Argon's 50. Google says it was built for long, multi-day engineering work, and that internal teams ran it for days on complex tasks like root cause analysis.
Access is the catch. The Verge reports Carbon is only available to a select group of cybersecurity defense companies for now, with expansion planned to other partners in Google's Fairwind Program. The Decoder reports it is going through the US government's voluntary pre-testing process because of its cyber capabilities. Pricing matches Argon at $2 per million input tokens and $10 per million output.
The independent numbers are mixed in useful ways. Per The Decoder, Artificial Analysis measured answer accuracy rising from 48% to 58% over Argon, while the hallucination rate climbed from 4% to 11%, still second lowest in that ranking. In agentic coding, Carbon trails Claude Code with Opus 5.5 by three points. It generates 2.3 times as many reasoning tokens as Argon, and Vals AI recorded Legal Research tasks averaging more than 50 minutes each.
My take
Benchmarks for unreleased models are a planning signal, not a build decision. For client work I care about three things this release shows:
- Longer thinking is not free. More reasoning tokens means slower runs, which matters inside a workflow that a customer is waiting on.
- Higher accuracy came with a higher hallucination rate. If a step must say "not found" instead of guessing, test that behavior specifically.
- Argon is the Gemini you can use today, and it is strong on real workflows. On Zapier's AutomationBench it leads at 51.3%.
The practical move is to keep the model a setting, not a rebuild. When Carbon opens up, run your own past cases through it and compare end states and cost before switching anything.
More posts
- Uber says AI agents booked four rides in its busiest week of October. Its app takes about 50 seconds to book oneNov 5, 2026
- A 2B parameter model beat frontier LLMs at pulling fields from invoices and IDs in its makers' benchmark, at 6 to 147 times lower costNov 5, 2026
- The best AI agent finished 29 of 45 real AWS tasks. About 80% of failures came from stopping too early, not from missing knowledgeNov 5, 2026
