Gemini 4 Carbon vs Opus 5.5: $16 per task, no launch date

Gemini 4 Carbon is live in benchmark results — not yet in production. SemiAnalysis’s new GDPx test, published Friday, puts Google’s latest frontier model at second place: just behind Anthropic’s Claude Opus 5.5, but well ahead on price. It scores $16.20 per task versus $51 for Opus 5.5 and $54.30 for OpenAI’s GPT-6 Astra.

That gap isn’t noise. It’s the difference between running a mid-scale inference workload on a $2,000 GPU cluster versus a $6,000 one, at scale, it reshapes margins. The benchmark covers 206 tasks across finance, materials science, legal reasoning, software engineering, and biology. Gemini 4 Carbon leads in cybersecurity tasks, outperforming both Opus 5.5 and GPT-6 Astra, though Meta’s Muse models beat it there and across alignment and safety-heavy subsets.

When those domains are removed, the ranking flips: GPT-6 Astra jumps to first, followed by GPT-6.1 Sol and Claude Sonnet 5.5, all ahead of Carbon. SemiAnalysis says that shift “better aligns with subjective impressions” of how each model behaves in real-world deployments where guardrails dominate output. Gemini 4 Argon, Google’s prior flagship, lands eighth, ahead of Meta’s Muse Wonder, but behind Wonder Full.