Gemini 4 Carbon's 4x cost claim has no source yet

Google prices Gemini 4 Carbon at $2 per million input tokens and $10 per million output tokens. Anthropic prices Claude Opus 5.5 at $4 and $20. Half the price is the most those two pages support, yet the coverage of the new GDPx benchmark says Carbon costs 4x less per task.

Verdict: unproven. SemiAnalysis's GDPx excerpt is marked "Paid" and truncated, carries no cost-per-task figure, and offers one cost sentence: "Opus wins the high end."

Half the price now, the same price later

Google's footnote on Carbon pricing reads: "After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply." The new price equals Opus 5.5's $4 and $20. The per-token comparison starts at half and ends at parity. It never reaches a quarter.

The introductory price also belongs to a model most readers have no way to buy. Google wrote: "Starting this week, we will make Carbon available to a subset of cyber defenders and will gradually phase access to additional trusted partners through our Fairwind Program." Google's October recap describes Carbon as in pre-release testing. I infer the $2/$10 figure is a launch price for a model not yet generally available.

The 12% mixes grading methods, and the 10.9% appears nowhere

SemiAnalysis published GDPx on Nov 02, 2026. It reports the best model scores 12%, and the excerpt never names which model. The coverage gives Opus 5.5 the 12% and Carbon 10.9%. The 10.9% appears in none of the five documents. Neither do the 206 professional tasks, the 14 partners or the corporate acquisition of nearly 4,000 files the coverage cites.

The coverage also describes strict pass-fail grading, and SemiAnalysis's method says otherwise. It reports pass@1 across three attempts, weights all domains equally, and grades with "a mix of binary and partial grading depending on the domain." Finance tasks are so difficult that all the models score 0% under binary grading, so the leaderboard reports average partial scores.

A 12% built this way blends binary and partial credit. The coverage's pass-fail wording overstates how uniform the grading is.

Per-token parity doesn't rule out a 4x per-task difference, but nothing published shows one

The strongest objection to everything above is this: absence from an excerpt is not absence from the report. The cost chart sits behind a paywall. Cost per task is tokens times price, and no document gives Carbon's tokens per task. A model at list-price parity with Opus 5.5 still costs far less per task if it burns far fewer tokens.

SemiAnalysis also follows its cost sentence with a flag: "the appearance of Gemini 4 Carbon is worth watching closely." Google's page carries a customer saying Gemini 4 Argon completed evals "at a quarter of the cost per task." Carbon launches at Argon's introductory price. A quarter of the cost is 4x. The pieces fit a 4x story.

They fit, and they prove nothing. The Argon quote comes from a customer's internal benchmark against unnamed peer models. It concerns neither GDPx, nor Carbon, nor Opus 5.5. "Worth watching" carries no measurement.

Cheaper per task also differs from more efficient. In GDPx's legal redlining task, 20 rollouts across four models produced mean partial scores of 44% for Opus 5.5, 25% for GPT-6 Astra, 2% for Gemini 4 Argon and 0.5% for GLM-5.3. I infer a model finishing less work costs less per task. Carbon was not among the four, so the result says nothing about Carbon itself. It shows what a cost figure without a score hides.

None of this closes the question. The chart might hold a 4x, and I have no access to it. Even if it does, GDPx ran each model in its native harness, Claude Code for Opus 5.5 and Antigravity for Carbon. No document says how cost per task was measured, including tokens, effort settings and harness differences.

Which Opus 5.5 you are comparing depends on the domains

SemiAnalysis says "Opus 5.5 is the best all-around model and wins in quant trading, hardware design, and chip design." Carbon "is the best bio model (likely due to weaker safety classifiers)" and edges out the win in software engineering. SemiAnalysis also says safety classifiers sandbag the overall scores of Opus 5.5 and GPT-6 Astra.

It then re-ranked without cyber, biology and alignment, and wrote that Carbon is "a real improvement from Google and the only model in the same ballpark as Opus and Astra." My reading is the ballpark claim belongs to the trimmed ranking. The coverage presents it as a general result.

Google's own table shows Carbon at 59 on AA Index v5.0 against 58 for Opus 5.5. Opus 5.5 leads on FrontierSWE 2 (62.3% against 61.2%) and OSWorld-2.1 (87.2% against 81.5%). GDPx is not in the table. Anthropic warns "benchmark margins have become a less reliable guide to real-world differences," then claims Opus 5.5 beats GPT-6 Astra on specific benchmarks "for about a fifth of the cost per task." Both vendors promote their own models' results on their own pages.

The one place a Carbon-versus-Opus multiple could live is the paid cost chart. Until someone quotes it, the only quarter-of-the-cost line in the record belongs to a customer talking about Argon.

Sources

• GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains. newsletter.semianalysis.com.
• Introducing Gemini 4 Carbon. blog.google.
• Gemini — Google DeepMind. deepmind.google.
• Introducing Claude Opus 5.5 \ Anthropic. anthropic.com.
• Google AI news: October 2026 updates. blog.google.
• Google's Gemini 4 Carbon Offers 4x Lower Cost Than Opus 5.5 on New AI Agent Benchmark. HuggingNews, 3 November 2026. The headline this piece checks.