Gemini 4 Carbon trails Claude Opus 5.5 in new benchmark but wins on cost

Anthropic's Claude Opus 5.5 tops SemiAnalysis's new GDPx frontier benchmark, but Google's Gemini 4 Carbon sits just behind it at roughly a third of the cost: $16 per task against Claude's $51.

SemiAnalysis, an AI research and analysis firm, published GDPx as a frontier-level benchmark covering 206 tasks across finance, materials science, cybersecurity, software engineering, legal, biology and several other domains. The leaderboard scores every major frontier model from OpenAI, Anthropic and Meta, listing both score and cost.

The headline result is close at the top. Claude Opus 5.5 leads, with Gemini 4 Carbon fractionally behind — a gap the report describes as relatively minimal. The economics separate them sharply. At roughly $16 per task, Gemini 4 Carbon also beats OpenAI's GPT-6 Astra, which scored lower while costing about $54 per task, more than both leaders.

The price advantage is the story here, but it comes with an asterisk the report itself supplies.

What happens when safety domains are removed?

The report notes that excluding cybersecurity, biology and alignment reshuffles the rankings. Safety guardrails (built-in refusals and restrictions in Claude and GPT models) put GPT-6 Astra at the top once those domains drop out, with GPT 6.1 Sol and Claude Sonnet 5.5 also outranking Gemini 4 Carbon.

SemiAnalysis says this adjusted ordering better aligns with their “subjective impressions” of each model. In plain words: once the safety-heavy tasks are stripped away, Google's near-win at the top disappears.

Where do Google's other models land?

Google's existing frontier model, Gemini 4 Argon, ranks eighth on both versions of the leaderboard. It beats Meta's Muse Wonder but sits below Wonder Full. Argon's costs in this benchmark run only slightly lower than Carbon's, which raises the question of why a customer would choose the older model at a near-identical price.

Both Google models outscore OpenAI and Anthropic on cybersecurity tasks specifically. Meta's Muse models beat them all, and the benchmark found Muse best in that domain “by a meaningful margin.”

When can customers actually use it?

Google announced Gemini 4 Carbon on October 28 and claimed “significant gains.” The public release date remains unconfirmed. Google says the model is “coming soon” for most customers — a phrase that carries no date and no commitment.

Google's own benchmark, published alongside the announcement, shows the new model beating GPT-6 Astra and Claude Opus 5.5 in many aspects. Vendor-run benchmarks warrant skepticism; SemiAnalysis's independent figures, which show Carbon second rather than first, are the more credible reference point.

For marketers and enterprises budgeting AI spend, the takeaway is concrete: near-frontier performance at roughly $16 per task versus $51 for the leader is a real cost differential — provided the safety-domain caveats and the missing release date don't shift the math.