Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- Gemini 4 Carbon scores 59 on Coding Agent Index, ranking second only to Claude Opus 5.5.
- Tasks cost $7.76 on average—23% cheaper than Opus 5.5 due to a 29% reduction in generated output tokens.
- Hallucination rate on AA-Omniscience v1.1 dropped to 11% compared to 50% for frontier peers.
- Prompt caching yields a 95% discount ($0.20 per 1M cached tokens).
Key facts
- Intelligence Index v5 Score
- 59 (Rank #1)
- Coding Agent Index Score
- 59 (Rank #2 behind Opus 5.5 at 62)
- Standard Pricing (Input / Output)
- $4.00 / $20.00 per 1M tokens (50% launch discount available)
- Cached Input Pricing
- $0.20 per 1M tokens (95% discount)
- Hallucination Rate (AA-Omniscience)
- 11% (vs 50% on Claude Opus 5.5)
Source: Key facts — Intelligence Index v5 Score
Source: Key facts. Intelligence Index v5 Score: 59 (Rank #1). Each figure is copied from the story.
- Intelligence Index v5 Score59 (Rank #1)
Performance and Benchmark Gains
Google's Gemini 4 Carbon has captured the #1 position on the Artificial Analysis Intelligence Index v5 with an aggregate score of 59. Carbon leads across four key evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience, and AutomationBench-AA. On the Coding Agent Index, running inside Google's Antigravity CLI, it scores 59, placing second directly behind Anthropic's Claude Opus 5.5 in Claude Code (62). It reaches 64% on Terminal-Bench 4.0 and 51% on FrontierCode v1.1.
Pricing and Token Economics
Carbon matches the standard pricing tier of Gemini 4 Argon at $4.00 per 1M input tokens and $20.00 per 1M output tokens, with Google offering an ongoing 50% launch discount for early users. Cached input tokens benefit from a 95% reduction down to $0.20 per 1M tokens. Due to conciseness gains—using 123k output tokens per task compared to 173k for Claude Opus 5.5—the average cost per index task is $7.76, roughly 23% lower than Opus 5.5 ($10.04).
Hallucination and Multimodal Capabilities
On the updated AA-Omniscience v1.1 evaluation, Carbon recorded a score of 54 with an accuracy of 58%. Its measured hallucination rate stands at 11%, outperforming Opus 5.5 and GPT-6 Astra (both measured at 50% on this benchmark). The model accepts multimodal inputs spanning text, images, video, and audio, and produces text output.
Try it in 2 minutes
# Example invoking Gemini Carbon with cached system instructions via curl
curl -X POST https://generativelanguage.googleapis.com/v1beta/models/gemini-4-carbon:generateContent \
-H "Content-Type: application/json" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-d '{
"contents": [{"parts": [{"text": "Analyze the repository structure and list test gaps."}]}],
"generationConfig": {"temperature": 0.2}
}'When to use
- High-throughput agentic workflows where output token costs and prompt caching make Claude Opus 5.5 too expensive.
- Tasks demanding scientific reasoning or low factual hallucination rates.
What to do today
- Check Google AI Studio or Vertex AI access to verify eligibility for the Gemini 4 Carbon rollout.
- Review agentic architectures to take advantage of the $0.20/1M cached token rate for repetitive system prompts.
- Compare Antigravity CLI output against Claude Code runs on your local test suites.
Step names are taken from What to do today.
Step names are taken from What to do today. Check Google AI Studio or Vertex AI access to verify eligibility for the Gemini 4 Carbon rollout.. Review agentic architectures to take advantage of the $0.20/1M cached token rate for repetitive system prompts.. Compare Antigravity CLI output against Claude Code runs on your local test suites.
- Check Google AI Studio or Vertex AI access to verify eligibility for the Gemini 4 Carbon rollout.
- Review agentic architectures to take advantage of the $0.20/1M cached token rate for repetitive system prompts.
- Compare Antigravity CLI output against Claude Code runs on your local test suites.
What the community says
It is vaporware at this stage. Also, their tooling still is complete shit. I'll not be forced to use Antigravity thing even if it is available and clearly better than Opus or Astra.
I similarly had Argon respond with a vast hallucination. It surprised me because it wasn't a little wrong on the details but a fully false verbose response that completely fabricated everything.